acknowledgementir.amu.ac.in/11964/1/t10316.pdf · 2018-11-22 · acknowledgement ii i express my...
TRANSCRIPT
i
Acknowledgement
"He who does not thank people, does not thank ALLAH" – Prophet Muhammad
(peace and blessings be upon him).
Unending glory is expedient to the Almighty, the Exalted; who granted me the
primary inspiration and stamina all along to complete this humble work. This small
contribution, if just and correct, is only a drop of appreciation for his Ocean of
munificence, without his blessings the work would not have seen the light of the day.
The completion of a research work has never been a ‘one man show’ but collective
efforts of all the well-wishers. This thesis has been very exciting and challenging for
me and I have been accompanied by a great number of people whose contributions are
worth to be acknowledged as Dwight Frindt has said, “Acknowledgment and
celebration are essential to fuelling passion making people feel valid and valuable
and giving the team real sense of progress and makes it all worthwhile.”
There are no proper words to convey my deep gratitude and respect for my thesis
supervisor and my mentor Prof. Jamshed Siddiqui, Professor and Chairman,
Department of Computer Science. I can never manifest the true sense of thanks to his
kind support which he has provided throughout the entire period of my Ph.D. I
acclaim his courage, respect his decisions, learn from his knowledge and reverence
the personality he has. The thesis would have not been accomplished successfully
without his kind support and sincere attention.
I am very much indebted to Co- Supervisor of thesis, Dr. Rashid Ali, Associate
Professor, Department of Computer Engineering, Aligarh Muslim University for
inspiring me to become an independent researcher and helped me realize the power of
critical reasoning. He also demonstrated what a brilliant and hard-working researcher
can accomplish. His advice on my research work has been invaluable.
My supervisors not only taught me to be a good scholar, but to be a good person in
the life also. I am really grateful to them for giving the Midas touch to my thesis.
Acknowledgement
ii
I express my sincere gratitude to Prof. Mohammad Ubaidullah Bokhari and Mr.
Suhel Mustajab, ex-Chairmen, Department of Computer Science, for providing me all
the necessary research facilities in the Department. I would like to convey my
heartfelt thanks to faculty members of the Department, Prof. Rafiqul Zaman Khan,
Mr. S. Maheshwari, Ms. P. Bala, Dr. Asim Zafar, Dr. Tamanna Siddiqui, Mr. Shahid
Masood, Dr. Arman Rasool Faridi, Ms. Sehba Masood, Mr. Faisal Anwer, Dr.
Swaleha, Dr. Sajid, and specially Dr. Mohammad Nadeem for their endorsements. I
also acknowledge my deepest gratitude to the staff of our lab.
I am immensely grateful to all members of research lab of the Department; Dr.
Nazir Ahmad, Dr. Suby Khanam, Dr. Yahya, Dr. Haider Khalaf Jabbar, Mr. Oqail
Ahmad, Kashif, Riyaz, Ausaf, Parvej, Rizwan, Shabbir, and Mr. Mayank Srivastava.
They generously gave their time to offer me valuable comments toward improving my
work. In particular Dr. Zaki Ahmad Khan and Mr. Faraz Hasan showed me the great
power through their innovative abilities.
It is delightful for me to thank my seniors Dr. Shadab Alam, Dr. Shamsh Tabrez
Siddiqui, Dr. Javed Ali, and Dr. Hattem for their appreciable advices.
It will be injustice, if I will not remember my dear friends at this juncture for being
with me in all my good and bad times during my long stay at the Aligarh Muslim
University. I feel very blessed when I count few names among them like Dr. Kamran
ahsan, Dr. Tanweer Khan, Dr. Tariq Sheikh, Dr. Imran Hussain, Mr. Humayun,
Ahmad Danish, Faheemuddin Malik, Fahim Akhtar, Saifullah, Shamsuddin and
Hasnain. I have been residing in the hall of residence for last 13 years and being loved
equally from the resident of respective hostels Mumtaz, Morison, Aftab and Mac
Donnell. My juniors at Hostel and Departments are worth mention here. To name a
few, Shuez, Faizan, Nafees, Faisal, Anas, Aziz, Zeyaullah, Farhan, Huzaifa, Ahmar,
Abdul Salam, Rashid, Abuzar, Adil, Haris, and my room partner Shanu.
Apart from the above, I would like to extend my heartfelt emotions to my
childhood friends Adnan Arif, Syed Tanweer, Shahab Akhtar and Meraj. I am also
thankful to Mobashir and Shadab for their technical support which helped me in
accomplishing my research work.
I deeply thank the other faculties of our University from different departments for
their support and kind attentions in my need. Specially, Prof Sufyan Beg has helped a
lot and I must mention him for his entire support during my UC, Berkeley visit for
Acknowledgement
iii
presenting my paper on an International Conference. Dr. Faiza Abbasi has encouraged
me and provided various platforms to boost my personality. Dr. Sagheeruddin, Prof.
Abdul Mannan, Dr. Obaidullah Khan and Professor Javed Arif have remained source
of encouragement and stimulation for me always.
Words cannot pay them back their kind attentions, care, guidance and love, I wish
for long age and health to the source of my spiritual power and inspirations, Prof.
Sanaullah Khan and Prof. Nadir Ali Khan, as they are light for me in the dark night.
I am deeply grateful to the University Grant Commission, New Delhi, for
providing me financial assistance in the form of “MANF” during my research, and
Vice Chancellor of the Aligarh Muslim University who worked extraordinary hard to
maintain peace in the campus and provided a more competitive and educational
environment in the University.
I pay my immense gratitude to my family members, whose Dua remain a source of
courage and inspiration for me ever, specially my big Brother Asif Sohail who has
sacrificed enough for this Degree. Also, my brothers, Mr. Akif, Mr. Absar, Mr.
Aehtasham, Jami and Sharjeel, brothers in law, Mr. Fakhre Alam and Nizamuddin,
my loving sisters, and sisters in law for their love, care and Dua.
Lastly, I would like to pay my sincere thanks and gratefulness to most influencing
character of my life, my mother; whose constant encouragement and support acted as
an impetus for working hard and completing the work with sincerity. There are no
words that can express gratitude for her love, affection and patience. She always stood
by my side, have faith in my work and always prayed for my success.
Shahab Saquib Sohail
List of Abbreviations
iv
Abstract
The proliferation of the Internet has attracted the masses for its wide application and
people have assimilated it in their lives. The huge inclination of the people towards
using Internet and other means of technology for various daily life activities like
online shopping has created problems of data overload. The growing data has
developed the problems for users in selecting the exact item of their preferences over
a large amount of available product. Recommender Systems (RS) come into the
picture to provide users a personalized and best suited item by saving the times in
searching the desired product, and reducing the complexities of using modern tools.
In the mid-90s, RS were introduced inclusively, prior to which these systems were
treated as an Information Retrieval (IR) approach. The Collaborative Filtering (CF) is
the most widely used recommendation technique along with other existing
recommendation approaches. The other well-known techniques which are used for
developing recommender systems are mainly, Content based (CB) or Reclusive
Methods (RM), Knowledge Based approach (KB), Demographic Filtering (DF),
Hybrid Approach (HA) and Context Aware approach for Recommender System
(CARS).
These techniques have been used by researchers and there is a significant increase
of researches in RS recently. The collaborative filtering (CF) based recommendation
approach tries to explore the recommendation from other customers whose choices
are similar to the target customers (i.e. customer for whom the recommendation is
made). Unlike collaborative filtering, reclusive approach finds similarities between
items without any collaboration of users. The recommender systems based on
demographic filtering also use similarity measures as a metric. But instead of finding
similar rated items by neighbor users, it tries to find the similarity between users’
demographic information like, age, sex, occupation, etc.
List of Abbreviations
v
The knowledge based system has been seen as a better substitute for above
discussed approaches. The idea which differentiates knowledge based systems from
other systems, is the degree of importance it gives to the following two domains –
a) user’s requirement
b) Characteristic of the recommended items.
The above area of expertise helps in achieving users’ satisfaction by fulfilling
their needs. Certainly, an approach for building recommender system which needs
either explicitly defined set of recommendation rules or some sort of similarity
measures from prior purchase history of the users is perceived as knowledge based
approach for recommender system.
There are several issues with the existing approaches. One of the severe
problems is the cold start problem. We propose a solution to cold start problem which
is based on the consensus ranking of the item that suits majority of the group to which
user belongs. However, for this approach we need to know similar-like user
surroundings. The approach may save time and ease the complexities involved in the
recommendation. In this work, we have considered problem of book recommendation
for computer science graduate students in Indian perspectives. Finding each user’s
preferences and providing personalized recommendation to all is time consuming and
extra efforts are required in it. Also, cold start issue will remain a threat forever. As a
solution, all graduate students of Indian Universities are considered as member of
same group. The top N books amongst these universities are obtained by observing –
a) what the best universities are recommending and, b) what the students have their
opinion about these books. By finding the best book with some experimented
suggestible approach, we may provide a good recommendation to large user without
unknown prior preference (UPP) problem. The results of the suggested approaches are
discussed in this work. The comparisons of the proposed approaches are made on the
basis of 8 different parameters. The parameters are P@10, FPR@10, FNR@10, Mean
Average Precision (MAP), Mean Reciprocal Rank (MRR), Mean Absolute Error
(MAE), Root Mean Square Error (RMSE), and Spearman rank correlation coefficient.
The proposed Opinion Mining Technique has performed well and produces a better
result for each parameter.
vi | P a g e
TABLE OF CONTENTS CONTENTS PAGE NO Certificates and Declarations Dedication Acknowledgement i Abstract iv Table of Contents vi List of Tables xii List of Figures xv List of Abbreviations xvii
Chapter 1: Introduction ..................................................................................................... 1
1.1 Recommender Systems: An Introduction .................................................................... 1
1.2 Application of Recommender Systems ......................................................................... 1
1.3 Problems and Issues in Recommendation Approaches.......................................... 2
1.3.1 Cold Start ................................................................................................................................. 3
1.3.2 Missing of Absolute Ranking ........................................................................................... 3
1.3.3 Personalization for Community Recommendation ................................................ 4
1.4 Web Mining Techniques ................................................................................................... 4
1.4.1 Web Mining Techniques as a Solution to the Existing Problems of Recommender Systems .................................................................................................................. 6
1.5 Organization of the Thesis ............................................................................................... 8
Chapter 2: An Overview of Recommender Systems ................................................ 11
2.1 Introduction: ....................................................................................................................... 11
2.2 Previous Review Studies ................................................................................................. 12
2.3 Types of Recommender Systems .................................................................................. 15
vii
2.3.1 Collaborative Filtering based Recommender Systems ........................................ 16
2.3.1.1 Item based and User based CF techniques ..................................................................18
2.3.1.1.1 Association rule mining between preferences of neighbor of users........19
2.3.1.1.2 Rating based recommendation .............................................................................19
2.3.1.1.3 Choice based recommendation ............................................................................20
2.3.1.1.4 Recommendation based on similarity in the users’ preferences for common items ..............................................................................................................................21
2.3.1.1.5 Tagging based recommendation ..........................................................................22
2.3.1.2 Model based CF techniques .............................................................................................22
2.3.2 Reclusive Methods based Recommender Systems................................................ 25
2.3.2.1 Heuristic based Reclusive Recommendation ..............................................................26
2.3.2.2 Model based Reclusive Recommendation ...................................................................27
2.3.2.3 Web Mining based Recommendation ..........................................................................28
2.3.3 Demographic Filtering based Recommender Systems ....................................... 30
2.3.4 Knowledge based Recommender Systems .............................................................. 32
2.3.4.1. Case-based Recommendation: .......................................................................................35
2.3.4.2. Constraint based Recommendation:.............................................................................36
2.3.5 Hybrid Recommender Systems .................................................................................... 37
2.3.5.1. Hybrid Recommender Systems based on Collaborative Filtering dominated Reclusive Method ...............................................................................................................................38
2.3.5.2. Hybrid Recommender Systems based on Reclusive Method dominated Collaborative Filtering Techniques ..............................................................................................40
2.3.5.3. Hybrid Recommender Systems based on unified Reclusive Method and Collaborative Filtering Techniques ..............................................................................................40
2.3.5.4. Hybrid Recommender Systems based on Subsequent Integration of separately applied Collaborative Filtering Techniques and Reclusive Method ..................................41
2.3.5.5. Hybrid Recommender Systems based on Integration of Collaborative Filtering and Reclusive Method with Knowledge based System ..........................................................42
2.3.5.6. Other Hybrid Recommender Systems using Collaborative Filtering Techniques ...........................................................................................................................................42
viii
2.3.5.7. Other Hybrid Recommender Systems using Reclusive Method ......................... 43
2.3.6 Context Aware Recommender Systems .................................................................... 44
2.3.7 Social Network based Recommender Systems....................................................... 48
2.3.8 Soft Computing Techniques based Recommender Systems ............................. 49
2.4 Summary .............................................................................................................................. 52
Chapter 3: Link Mining based Book Recommendation Approach ...................... 54
3.1 Introduction ........................................................................................................................ 54
3.2 Book Recommendation using Positional Aggregation based Scoring Technique .................................................................................................................................... 56
3.2.1 Positional Aggregation based Scoring Technique ................................................ 57
3.2.2 Book Recommendation Approach using Positional Aggregation Scoring . 57
3.3 Results and Discussions .................................................................................................. 63
3.3.1 Dataset ................................................................................................................................... 63
3.3.1.1 Selection of top Universities ............................................................................................ 63
3.3.1.2 Courses included from top Universities ...................................................................... 64
3.3.1.3 Prescribed books by top Universities: ........................................................................... 65
3.3.2 Experimental Results ....................................................................................................... 66
3.4 Summary .............................................................................................................................. 71
Chapter 4: Book Recommendation based on Soft Computing Approaches ...... 72
4.1 Introduction ........................................................................................................................ 72
4.2 Ordered Weighted Aggregation .................................................................................. 74
4.3 Book Recommendation based on Ordered Weighted Aggregation ............... 76
4.4 Book Recommendation Approach using Ordered Ranked Weighted Aggregation ................................................................................................................................. 77
4.4.1 Ordered Ranked Weighted Aggregation ................................................................. 77
ix
4.4.2 Book Recommendation based on Ordered Ranked Weighted Aggregation ............................................................................................................................................................. 79
4.5 Results and Discussions ...................................................................................................81
4.5.1 Dataset ................................................................................................................................... 81
4.5.2 Experimental Results ....................................................................................................... 81
4.6 Summary ...............................................................................................................................91
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation ............................................................................................................... 93
5.1 Introduction: ........................................................................................................................93
5.2 Customer Reviews..............................................................................................................95
5.2.1 Issues while handling Online Reviews: .................................................................... 95
5.3 Feature Extraction and Selection .................................................................................97
5.4 Scoring Technique for Extracted Feature .............................................................. 102
5.4.1 Opinion Score Calculation ......................................................................................... 102
5.4.1.1 Positive words: ................................................................................................................... 103
5.4.1.2 Negative words: ................................................................................................................ 103
5.4.1.3 Reciprocal terms ............................................................................................................... 105
5.4.1.4 Highly expressible words .............................................................................................. 107
5.4.2 Weight Assignment to Features ................................................................................ 108
5.5 Results and Discussions ................................................................................................ 109
5.6 Summary ............................................................................................................................ 113
Chapter 6: Evaluation of Recommender Systems .................................................. 115
6.1 Introduction: ..................................................................................................................... 115
6.2 Previous Evaluation Studies ........................................................................................ 118
6.3 Evaluation Metrics .......................................................................................................... 119
x
6.3.1 P@10 .................................................................................................................................. 120
6.3.2 FPR@10 ............................................................................................................................. 121
6.3.3 FNR@10............................................................................................................................. 121
6.3.4 Mean Average Precision.............................................................................................. 121
6.3.5 Mean Absolute Error .................................................................................................... 121
6.3.6 Mean Reciprocal Rank ................................................................................................. 121
6.3.7 Root Mean Square Error ............................................................................................. 122
6.3.8 Spearman rank Correlation Coefficient ............................................................... 122
6.3.9 Modified Spearman rank Correlation Coefficient ............................................ 123
6.4 Evaluation based on Experts’ Ranking using Explicit Feedback .................. 123
6.4.1 Evaluation Results based on Different Evaluation Metrics ........................... 125
6.4.1.1 Evaluation Results using Explicit Feedback based on Root Mean Square Error .............................................................................................................................................................. 126
6.4.1.2 Evaluation Results using Explicit Feedback based on Mean Absolute Error 128
6.4.1.3 Evaluation Results using Explicit Feedback based on P@10 ............................. 130
6.4.1.4 Evaluation Results using Explicit Feedback based on Mean Average Precision .............................................................................................................................................................. 132
6.4.1.5 Evaluation Results using Explicit Feedback based on FPR@10 ........................ 133
6.4.1.6 Evaluation Results using Explicit Feedback based on FNR@10 ....................... 134
6.4.1.7 Evaluation Results using Explicit Feedback based on Modified Spearman Rank Correlation Coefficient ................................................................................................................. 136
6.4.1.8 Evaluation Results using Explicit Feedback based on Mean Reciprocal Rank .............................................................................................................................................................. 138
6.4.2 Comprehensive Evaluation Measure ..................................................................... 139
6.5 Evaluation based on Implicit User Feedback ....................................................... 140
6.6 Architecture for Evaluation Scheme based on Implicit Feedback ............... 141
6.6.1 Vector Component of User feedback ..................................................................... 143
6.6.2 User Feedback based Scoring of Products ............................................................ 144
xi
6.6.3 User’s Sincerity Measure ............................................................................................ 147
6.6.4 Product Preference Score ............................................................................................ 150
6.6.5 User Personalized Ranking ........................................................................................ 152
6.7 Results and Discussions ................................................................................................ 153
6.7.1 Mean Reciprocal Rank obtained using Comprehensive Approach ........... 154
6.7.2 Precision@10 obtained using Comprehensive Approach ............................ 155
6.7.3 Mean Average Precision obtained using Comprehensive Approach ........ 156
6.7.4 FPR@10 obtained using Comprehensive Approach ........................................ 157
6.7.5 FNR@10 obtained using Comprehensive Approach ....................................... 159
6.7.6 Spearman Correlation value using Comprehensive Approach ................... 160
6.8 Relative Performance of the Recommender System using Proposed Comprehensive Approach and other Existing Evaluation Approaches ............. 161
6.8.1 Comparison of Proposed Comprehensive Approach with Existing Evaluation Strategies ................................................................................................................ 165
6.9 Summary ............................................................................................................................ 167
Chapter 7: Conclusion and Future Direction ......................................................... 169
7.1 Introduction ...................................................................................................................... 169
7.2 Conclusion ......................................................................................................................... 169
7.2 Future Directions ............................................................................................................ 172
References: ........................................................................................................................ 174
List of Publications : …………………………………………………….….200
xii
LIST OF TABLES
Table 1. 1: A summary of results contained in this thesis .................................................................. 10 Table 2. 1:A glance of the review studies on Recommender Systems ............................................. 14 Table 2. 2 : Collaborative Approach illustration.................................................................................... 16 Table 2. 3: Recommender Systems, Categories and Techniques ....................................................... 50 Table 3. 1: Top 4 ranked books by 5 universities .................................................................................. 58 Table 3. 2: : Conversion of Rank into Scores ........................................................................................... 58 Table 3. 3: : Pairwise comparison of books .............................................................................................. 59 Table 3. 4: Normalized preference score of books ................................................................................ 59 Table 3. 5: Positional Aggregated scores of books ................................................................................. 62 Table 3. 6: Preference score of books......................................................................................................... 62 Table 3. 7: Ranked books based on Positional Aggregation Scoring technique .......................... 62 Table 3. 8: Top 7 Indian Universities in QS ranking [244] ................................................................ 64 Table 3. 9: Syllabus of various Courses, offered at top Universities. ............................................... 65 Table 3. 10: Total number of books in the syllabus of corresponding courses in top
Universities .............................................................................................................................................. 66 Table 3. 11: Code and details for books on Compiler Design ........................................................... 67 Table 3. 12: Ranked list of book ‘compiler design’ by top universities .......................................... 68 Table 3. 13: Compiler design ranked books by top 7 Universities .................................................. 68 Table 3. 14: Rank to Score conversion of book Compiler Design .................................................... 69 Table 3. 15: Positional Score for book Compiler Design ..................................................................... 69 Table 3. 16: Ranking of book ‘compiler design’ using Positional Aggregation Scoring ........... 69 Table 3. 17: PAS based Ranking of different books ............................................................................... 70 Table 4. 1: Ranked books using relative quantifier most .................................................................... 76 Table 4. 2: Ranked books using relative quantifier As many as possible ....................................... 77 Table 4. 3: Ranked books using relative quantifier At least half ...................................................... 77 Table 4. 4: Ranked books based on Ordered Ranked Weighted Aggregation technique for
example 3.1 ............................................................................................................................................. 81 Table 4. 5: Code and details for books on Compiler Design .............................................................. 82 Table 4. 6: Ranked list of book ‘compiler design’ by top universities ............................................. 83 Table 4. 7: Compiler design ranked books by top 7 Universities ..................................................... 84 Table 4. 8: Rank to Score conversion of book Compiler Design ....................................................... 84
xiii
Table 4. 9: Positional Score for book Compiler Design ........................................................................85 Table 4. 10: Score obtained by recommendation approaches for compiler design ...................85 Table 4. 11: Five different ranking of book ‘compiler design’ ..........................................................86 Table 4. 12: Five different ranking of book ‘Discrete Mathematics’ ...............................................86 Table 4. 13: Five different ranking of book ’Artificial Intelligence’ ................................................87 Table 4. 14: Five different ranking of book ‘Data Structure’ .............................................................87 Table 4. 15: Five different ranking of book ‘Principal of Data Base’ ...............................................87 Table 4. 16: Five different ranking of book ‘‘Computer Graphics’ ..................................................88 Table 4. 17: Five different ranking of book ‘Software Engineering’ ...............................................88 Table 4. 18: Five different ranking of book ‘‘Operating System’ ......................................................88 Table 4. 19: Five different ranking of book Computer Network’ .....................................................89 Table 4. 20: Five different ranking of book ‘Theory of Computation’ ............................................89 Table 5.1: Features and related review terms ...................................................................................... 101 Table 5.2: Precision of extracted features ............................................................................................. 110 Table 5.3: weights distribution of features ........................................................................................... 111 Table 5.4: Score calculation example ..................................................................................................... 111 Table 5.5: Example of Final Score Calculation ................................................................................... 112 Table 5.6: Top 10 ranked books of all the courses using Opinion Mining Technique .......... 112 Table 6. 1: Root Mean Square Error of all books by different approaches................................. 126 Table 6.2: Mean Absolute Error of all books for different approaches ....................................... 129 Table 6.3: P@10 for all approaches ........................................................................................................ 131 Table 6.4: Mean Average Precision of different approaches. ......................................................... 132 Table 6. 5: FPR@10 for all techniques. .................................................................................................. 134 Table 6.6: FNR@10 of all books................................................................................................................ 135 Table 6.7: Modified Spearman Rank Correlation Coefficient by different approaches ......... 137 Table 6.8: Mean Reciprocal Rank of all techniques for different Courses ................................. 138 Table 6.9: Final values of parameters used to find error .................................................................. 139 Table 6.10: Final values of parameters used to find precisions and correlation ...................... 140 Table 6.11: Comprehensive evaluation measure ................................................................................ 140 Table 6.12: Illustration for the calculation of Normalized Products Importance Score ‘δ’ for
user 1. ..................................................................................................................................................... 146 Table 6.13: Correlation values of different products of Laptop ..................................................... 148 Table 6.14: Correlation values of different products of Printer .................................................... 148 Table 6.15: Correlation values of different products of Head Phone........................................... 149 Table 6.16: Correlation values of different products of Tablet ...................................................... 149
xiv
Table 6.17: Correlation values of different products of Smart Phone ......................................... 149 Table 6.18: List of users which are excluded after user’s sincerity analysis ............................. 150 Table 6.19: Criteria of preference for a product to be preferred by a user ............................... 151 Table 6.20: Normalized Products Importance Score for Laptop ................................................... 152 Table 6.21: Ranking of laptop by different users based on product preference score .......... 152 Table 6.22: Mean Reciprocal Rank of first ranked product of different items ......................... 154 Table 6.23: values of precision at k, for different items ................................................................... 155 Table 6.24: Mean Average Precision for different products ........................................................... 156 Table 6.25: values of FPR@10 for different products ....................................................................... 157 Table 6.26: Avg. FPR@5 and Avg. FPR@10 for all the items .......................................................... 158 Table 6.27: Values of FNR@10 for different products ..................................................................... 159 Table 6. 28: Avg. FNR@5 and Avg. FNR@10 for all the items ....................................................... 159 Table 6.29: Spearman correlation coefficient for different products .......................................... 160 Table 6.30: Mean Reciprocal Rank, P@5, Mean Average Preciison and Spearman Correlation
Coefficient for different approaches ............................................................................................ 162 Table 6.31: FPR@5 and FNR@5 for different approaches .............................................................. 164 Table 6. 32: Comparison of proposed Comprehensive Approach with existing evaluation
strategies ................................................................................................................................................ 166
xv
LIST OF FIGURES
Figure 2. 1: Collaborative Filtering Approach ........................................................................................17 Figure 2. 2: Reclusive Approach for Recommendation ........................................................................26 Figure 2. 3: Demographic Filtering based Recommendation Approach ........................................31 Figure 2. 4: Context Aware Recommender Systems Overview .........................................................45 Figure 2. 5: Example for Context Aware Recommender Systems using season based clothes
......................................................................................................................................................................45 Figure 3.1: An overview of link mining approach ................................................................................55 Figure 3. 2: Positional Aggregation Scoring based Book Recommendation System ...................61 Figure 4.1: Most Quantifier ..........................................................................................................................75 Figure 4. 2: As many as possible quantifier .............................................................................................75 Figure 4. 3: At least half quantifier .............................................................................................................76 Figure 4.4: Ordered Ranked Weighted Aggregation based Book Recommendation System...80 Figure 5.1: Demonstration of a review in Spanish ................................................................................96 Figure 5.2: Demonstration of a review in Russian ................................................................................96 Figure 5.3: Demonstration of a review in Portuguese .........................................................................96 Figure 5.4: Demonstration of a review in Greek ...................................................................................96 Figure 5.5: Screenshot displaying no reviews .........................................................................................96 Figure 5.6: Architecture of Book Recommendation using meta searching ...................................97 Figure 5.7: Screenshot of Seacrh Engine Result Page for books on Artificial Intelligence .......99 Figure 5.8: Customer review expressing the views about content ..................................................99 Figure 5.9: Review example of ‘understandability’ feature. ............................................................ 100 Figure 5.10: Review representing importance of physical attributes .......................................... 101 Figure 5.11: Review representing importance of Price .................................................................... 101 Figure 5.12: Precision of Extracted Features ........................................................................................ 110 Figure 6. 1: Block diagram for Evaluation of Book Recommendation Approaches ................ 125 Figure 6. 2: Average Root Mean Square Error for all techniques ................................................. 128 Figure 6. 3: Average of Mean Absolute Error of all the books for different quantifiers ........ 130 Figure 6. 4: Average P@10 for all books ............................................................................................... 132 Figure 6. 5: Mean Average Precision of different approaches ....................................................... 133 Figure 6. 6: Average FPR@10 for all books using different book recommender approaches
................................................................................................................................................................... 134 Figure 6. 7: Average FPR@10 for all books using different book recommender approaches
................................................................................................................................................................... 136 Figure 6. 8: Average of Modified Spearman rank correlation coefficient .................................. 137 Figure 6. 9: Average Mean Reciprocal Rank of all the books for different techniques .......... 139
xvi
Figure 6. 10: Block diagram of implicit User Feedback based Evaluation of Recommender Systems ................................................................................................................................................... 142
Figure 6. 11: Mean Reciprocal Rank of top rank-position for respective items using Comprehensive Approach. .............................................................................................................. 155
Figure 6. 12: P@k for different items using Comprehensive Approach ..................................... 156 Figure 6. 13: Mean Average Precision using Comprehensive Approach ................................... 156 Figure 6. 14: Average FPR@5 using Comprehensive Approach ................................................... 158 Figure 6. 15: Average FPR@10 using Comprehensive Approach ................................................. 158 Figure 6. 16: Average FNR@5 using Comprehensive Approach .................................................. 160 Figure 6. 17: Average FNR@10 using Comprehensive Approach ................................................ 160 Figure 6. 18: Spearman correlation coefficient between system ranking and Comprehensive
Approach based ranking .................................................................................................................. 161 Figure 6. 19: Comprehensive Veracity Measure of different approaches ................................. 162 Figure 6. 20: Average FPR@5 for all items ........................................................................................... 164 Figure 6. 21: Average FNR@5 for all items .......................................................................................... 165
xvii
LIST OF ABBREVIATIONS
Acronym Full form
CARS Context Aware Recommender Systems
CB Content Based
CEM Comprehensive Evaluation Measure
CF Collaborative Filtering
DF Demographic Filtering
FNR False Negative Rate
FPR False Positive Rate
HA Hybrid Approach
HRS Hybrid Recommender Systems
KBS Knowledge based Systems
MAE Mean Absolute Error
MAP Mean Average Precision
OMT Opinion Mining Technique
ORWA Ordered Ranked Weighted Aggregation
OWA Ordered Weighted Aggregation
P@k Precision at top K position
PAS Positional Aggregation based Scoring
RMSE Root Mean Square Error
SMRCC
Spearman Rank Correlation Coefficient
1 | P a g e
Chapter 1
Introduction
1.1 Recommender Systems: An Introduction
"Necessity is the mother of invention," the famous proverb is practically experienced
in our daily life as we are growing and moving towards the advancement in
technologies. These technologies are giving birth to the modern tools and techniques
to fulfill our daily needs. Today a huge number of users are using the Internet. The
developed countries like Germany and U.K have approximately 83% Internet users of
their population, whereas China leads the overall contribution to the Internet users in
the world, which counts to 22.4%. USA has 78.1% Internet users of their population,
a contribution of 10.2% of overall users in the world [1]. This accelerated increase in
the use of the Internet in recent days has changed the style people live, they think and
they work.
With the changing trends in technologies, daily life of an individual has also
changed at a very fast pace. People prefer online shopping for their needs more and
more. To make online shopping easy and reliable a good number of product
recommendation techniques are proposed by many researchers in recent decay[2], [3],
[4]. Recommender systems (RS) try to identify the need and preferences of users,
filter the huge collection of data accordingly and present the best suited option before
the users by using some well-defined mechanism.
There are several known and frequent used techniques to recommend products
including Collaborative Filtering (CF),Content based (CB) or Reclusive Methods
(RM), Knowledge Based approach (KB), Demographic Filtering (DF), Hybrid
Approach (HA) and Context Aware approach for Recommender System (CARS), etc.
1.2 Application of Recommender Systems
The popularity of the recommender system is evident from the fast and vast
development of these systems for various applications. In the present time where
researchers deal with big data, it is observed that the prime focus of recommender
Chapter 1: Introduction
__________________________________________________________________________________
2 | P a g e
system‟s research is its application study. The application area of the RS has spread
over various domain of the daily life which includes –
Academics
Business
Entertainment
Health and care
Sports
In academics, RS have been widely used for Book recommendations[5],[6], [7]
Institute recommendations[8], Research papers recommendations [9], [10],
Conference recommendation [11] and Courses recommendations[12], [13], etc. The
recommender systems have gained much popularity in providing a means of
entertainment due to early music and movie recommendations [14], [15], [16]. The
sub domains of entertainment [17], [18], where RS have been used include Music,
Movie, TV program, and Tourism [19]–[21], etc.
The application of recommender system for health has grown rapidly due to the
increase in the demand of health information systems. Various health related
recommendation methods have been proposed [22], [23], [24], [25], [26].
There are numerous recommender systems have been developed for various e-
commerce applications. These applications oriented RS help business users to get
information about products and services of other products. Online shopping
interchangeably termed as e-shopping has achieved a high rate of growth in recent
days. Many portals have been designed for e-commerce applications based on e-
shopping [27], [28], [29], [30].
1.3 Problems and Issues in Recommendation Approaches
The different techniques which are used in designing RS have their own advantages
and limitations. The collaborative filtering, reclusive approach and demographic
filtering are basically learning-based techniques. Somehow, all these techniques
exhibit cold start problems. We explain and illustrate the major issue with these
leading recommender system technology supported by suitable examples in the
followed section, along with other issues exist with other existing techniques.
Chapter 1: Introduction
__________________________________________________________________________________
3 | P a g e
1.3.1 Cold Start
Cold start problem occurs for new user as well as for new items. We call it unknown
prior preference (UPP) problem. Although, reclusive approach does not need other
user‟s preferences and purchase details, it can recommend best match to user‟s
preferences only if it knew how and what has been rated by users previously. Problem
arises when a new user comes to shop and the system due to lack of its past
experience fails to recommend an item that matches to its choice. When a new user
starts buying something, the reclusive approach fails to identify what to recommend?
Similarly, if a user has been purchasing few specific types of products from a
merchandiser, say clothes, if s/he starts to seek electronic gadgets, reclusive approach
alone is unable to make any better recommendation.
Collaborative approach needs a good amount of rating from neighbors of a user;
also it requires a good amount of rating for identifying neighbor users. Therefore,
newly launched items and items those do not have good number of rating have found
weaker recommendation, though how best it may suite to user. For example, if a
person who lives in a cool place, like Paris or New York, his/her choice would always
be warm fabrics while shopping clothes. The person would move to a relatively hot
place, say Chennai or Mumbai (India) for any business or tour purpose, it would be
difficult to recommend cotton clothes to the user against the profile already have been
established which have high ratings for warm clothe only. Thus, in C.F approach,
UPP problem remain usual. In the same way for a new user C.F approach lacks a
better recommendation. The same issue seems to happen for demographic filtering
based recommendation also.
Although the researchers have suggested knowledge based approach as a solution
to the cold start problem. In knowledge based approach knowledge engineer learns
the user‟s immediate requirement and matches it to product features without any
historical data of the product and users. The knowledge based approach does not rely
upon any prior information of users. However, finding dedicated users who must be
willing to devote a good amount of time so that system can learn their preferences, is
difficult.
1.3.2 Missing of Absolute Ranking
Usually, the C.F approach needs rating. The merchandiser and users have their own
rating scale and own perception and understanding for rating scale. Hence, there is no
Chapter 1: Introduction
__________________________________________________________________________________
4 | P a g e
standard rating parameter and lack of rating standard affects the recommendation
badly. Some of the sites use 5 rating scale whereas others have 10 rating scales. So if
we consider only how much star has been awarded by the user for a particular item, it
would be confusing. A product is rated 6 out of 10, and another is rated 4 out of 5, it
is obvious the latter is best rated but the system which only asks number of rating or
rating points, it would opt for former. Thus we need some aggregation operator that
can fit the difference in one view. Also, rating scale gives a relative preference idea
and not absolute ranking. Hence, a user while reading a book gives 4 stars on amazon
while the quotes from the user is indicating that the book is not up to the mark,
however is well. That is 4 stars for a user means different from other. Different users
perceive rating differently. Sometime a user gives 3 star to the best books he has ever
been read. Whereas another user might have rated 4 stars to an average book, thus,
there is a need of some operator that can eliminate the difference and project the
rating absolutely and not relatively.
1.3.3 Personalization for Community Recommendation
Generally the leading RS filtering techniques like CF and RM favor the philosophy of
personalization. Though personalized recommendation seems great idea while
predicting an item to a particular user, however, making recommendation for users
which belong to more or less same group and have similar requirements causes extra
time and effort and repetition of the same process for different users. Let us consider
the book recommendation problem for students of same course and year. The
variations in the preferences of the students are possible but the need of these students
is equally important and similar. Also, there are specified domains of collection of
books for the same courses students. Thus, instead of applying personalized
recommendation approach it seems adequate to make use of a group recommendation
technology and same techniques and single ranked recommendation can be one
answer to several simultaneous queries [31].
1.4 Web Mining Techniques
Web mining is the application of data mining techniques to extract knowledge from
the Web data [32]. The researchers in this work have proposed web mining techniques
as a solution to various problems with the existing recommender systems‟
technologies which have been discussed above. In this work, web mining techniques
Chapter 1: Introduction
__________________________________________________________________________________
5 | P a g e
have been utilized for the recommendation of several electronic products and books.
Web mining techniques include – a) web usage mining, b) web structure mining and
c)web content mining.
Web usage mining makes use of log information of users on web and according to
their activity the items which match to their preferences are recommended to users.
Web structure mining, also referred as „link mining‟, incorporates the link available in
a web page and explores those links to find the best options for a user. Link mining is
a new challenging area where statistical modeling is performed for relational learning
[33]. It gives the importance of a website by finding out the backward links and
forward links. A forward link of a web page „A‟ of any website is the
recommendation of different web pages by this particular page, whereas the backward
link gives that by how many different pages this particular page is recommended, i.e.
linked. The importance of a page is determined by the number of backward link it has.
If the backward link is more, a page‟s importance is higher. For a high value page, its
entire forward links are considered to be highly valuable. Keeping the above concept
in consideration, we have chosen top universities amongst the Indian universities and
checked their recommendation for different courses of computer science, it‟s evident
that recommendation of a book by a high class university will eventually increase the
importance of the recommended books.
Web content mining is a process of extracting useful information from content of
the web. These contents may be the details of the items available on the web and
opinion of the users, etc. As far as online purchasing of products is concerned;
opinions from the customers are seen as a base to analyze the features of a product
and assess the requirement of a user. Customers' reviews are the basis for opinion
mining technique. Finding and summarizing the opinion from huge amount of
customers reviews, is also very tedious for business. The summary of reviews is
worth for the job.
For researchers, Opinion mining is a very hot topic in the field of data mining. The
main issue to consider is to find (a) product feature and (b) analysis comments,
whether positive or negative, as described in [5, 6, 7]. People generally use to analyze
some pre-determined terms to interpret it as a positive or negative comments, like,
better, good, nice, well written, highly recommend etc. are treated as positive terms
and worst, time consuming, bad, not recommended, etc. are termed as negative
Chapter 1: Introduction
__________________________________________________________________________________
6 | P a g e
comments. In [12], opinion retrieval is perceived as a two-step task, finding relevant
documents and re-ranking these documents by opinion scores. The reviews are given
by the human and it is very evident that to understand the review one should perceive
it as human being. Finding the comments is not sufficient. Sometimes things are
different then what they seem to appear. Let us consider the following example:
"I highly recommend this book for those who want to waste their time and money.
If you are really sincere to get some knowledge into your bucket, another one is the
better option"
Though the sentence above has terms like highly recommend and better but both
terms are being used in a negative sense for some specific book, keeping only the
positive and negative aspects of the terms and processing on these basis is not
sufficient alone to extract opinion for a better conclusion.
1.4.1 Web Mining Techniques as a Solution to the Existing Problems of
Recommender Systems
In this work, the researchers have tried to employ all the three discussed web
mining techniques for designing the recommendation methodology. The details of
mathematical verification of the procedure are elaborated in the respective chapters of
the thesis. Here, we have recommended top books on different disciplines of
computer science by using different web mining techniques. For opinion mining; we
have suggested various algorithms which consider the above mentioned issues in
finding the orientation of users in their opinions or reviews. For web structure and
web usage mining; we have tried to weight the importance of most valuable
universities (which may be consider as valuable links in link/structure mining) for
recommendation and validation recommender systems are performed on the basis of
user‟s behavior towards reviews of a product (which represents mining of the web
usage).
We propose a solution to cold start problem which is based on the consensus
ranking of the item that suits majority of the group to which user belongs. However,
there are two important aspects in it, first, to know similar-like user surroundings and
second it may not be personalized recommendation. However, this approach may save
time and ease the complexities involved in the recommendation. Let us consider book
recommendation problem graduate students of any university. Finding each user‟s
preferences and providing personalized recommendation all is time consuming and
Chapter 1: Introduction
__________________________________________________________________________________
7 | P a g e
efforts are required in it. Also, cold start issue will remain a threat forever. As a
solution, all the graduate students of same course in a University can be considered as
member of one group. Top N books amongst several universities can be obtained by
observing – a) what the best universities are recommending and, b) what the students
have their opinion about these books. By finding the best book with some
experimented suggestible approach, a good recommendation can be provided to large
user without unknown prior preference (UPP) problem. We have tried to incorporate
the above approach for the recommendation of books. The explanations and detail
procedure for different techniques used is elaborated in consecutive chapters.
Ordered Weighted Aggregation (OWA) and Ordered Ranked Weighted
Aggregation (ORWA) are used to include absolute ranking, as discussed in section
1.3.2 that missing of absolute ranking is a major concern. These techniques help in
aggregating the users‟ heterogeneous ranking to obtain a final aggregated result. We
have tried to utilize OWA and incorporated the proposed ORWA for aggregation
purpose.
With the above discussions in the considerations, we have also suggested a ranked
recommendation approach for books which aggregates the several ranking of the top
universities (which is considered as authorities) and employ link mining approach in
the recommendation process. On the one hand it handles the cold start issues and on
the other hand it eases the complexities of personalized recommendation to huge
number of users and replaces it with a single ranked recommendation. Chapter 3 and
Chapter 4 deals with the above issues and a comprehensive approach based on the soft
computing and link mining approaches will be discussed.
Opinion mining avoids user‟s rating and rather it emphasizes on user‟s reviews.
Thus, opinion mining can be a good solution to deal with the issues those have been
encountered with rating based recommendations. The issue with the opinion mining
which has been discussed in section 1.4 has not been adequately addressed yet. We
have tried to propose the solution for these problems. Several algorithms will be
discussed in Chapter 5 which we believe would overcome the prevailing issues to an
extent.
Chapter 1: Introduction
__________________________________________________________________________________
8 | P a g e
1.5 Organization of the Thesis
The thesis contains seven chapters. In Chapter 1, the introduction of the thesis is
presented which gives an overview of the problem formulated and proposed solutions.
Chapter 2 reviews extensively the literature and presents different categories possible.
Chapter 3 and Chapter 4 presents link mining and soft computing approaches
respectively for the recommendation of items in general, and books in particular.
Chapter 5 gives feature based recommendation of books which exploits opinion
mining techniques. In Chapter 6, evaluation of recommender systems based on
explicit as well as implicit feedback is discussed. The need of explicit and implicit
feedback is also discussed in details. Finally we have concluded the complete work in
Chapter 7. The Chapter-wise description is as follows;
Chapter 1: The recommender systems, its need and applications are introduced. A
highlight on the issues encountered in existing techniques is presented. The proposed
techniques which could be helpful in solving the emergent issues are briefly
described. Also, the main idea contained in this research work is featured in this
section.
Chapter 2: Review of Recommender systems have been performed in Chapter 2. The
literature survey is carried out by studying more than 200 recent research papers,
published in reputed conferences and peer reviewed journals, on the topic. The
various limitations and shortcomings of the existing techniques have been mentioned.
Chapter 3: Chapter 3 deals with an extensive research work supported by several
architectures designed for proposed link mining techniques which also incorporates
positional aggregation to aggregate the top universities recommended books and
provide students with the best books for their syllabus. The results of
recommendations made by the proposed approaches are shown at the end of the
chapter. However, the comparison of the approaches is explicitly discussed in Chapter
6.
Chapter 4: In Chapter 4, Soft Computing techniques have been used, in addition to
this, the fuzzy aggregation have been incorporated for the aggregation purpose. The
results of recommendations made by the proposed approaches are shown at the end of
Chapter 1: Introduction
__________________________________________________________________________________
9 | P a g e
the chapter. The comparison of the approaches with positional aggregation is
explicitly discussed in Chapter 6.
Chapter 5: In this chapter, the main consideration is to introduce opinion mining for
product recommendations. We have taken books as a product here. The
recommendation of books is discussed with the help of opinion extraction and feature
selection algorithm. These algorithms are designed by keeping in view those
considerations which are not well studied in the literature for the recommender
systems.
Chapter 6: A Comprehensive Evaluation Approach has been discussed in Chapter 6.
The comparison of the proposed approaches are discussed which is based on Explicit
feedback. The comparisons are made on the basis of eight (8) different parameters.
The parameters are P@10, FPR@10, FNR@10, Mean Average Precision (MAP),
Mean Reciprocal Rank (MRR), Mean Absolute Error (MAE), Root Mean Square
Error (RMSE), and Spearman rank order correlation coefficient. An additional
Comprehensive evaluation approach for implicit user feedback is also suggested and
an existing RS has been evaluated using proposed approach. The evaluation approach
may help in assessing the performance of any recommender systems as well. The
suggested approach uses implicit user feedback and recommends only those products
which are preferred by the users.
Chapter 7: It concludes the overall work and emphasizes our contributions in the
research work carried out in the thesis. It also focuses on the scope for future
extension of the work.
The main contributions in the thesis has been summarized and presented in Table 1.1.
Chapter 1: Introduction
__________________________________________________________________________________
10 | P a g e
Table 1.1: A summary of results contained in this thesis
Serial No.
Recommendation
based on Web Mining
Techniques
Contribution Chapter Publications
1.
Survey of the existing
literature
Literature Survey of RS
from Ecommerce
perspective
2 [28], [32]
2.
Link Mining based
Book Recommender
Systems
Incorporated Positional
aggregation technique for
the recommendation of
books.
3 [34]
3.
Soft Computing based
Book Recommender
Systems
ORWA is proposed and
utilized for the
recommendation of
books.
4 [5]
4.
Fuzzy technique based
Book Recommender
Systems
OWA is employed for
book recommendation 4 [35], [36]
5.
Opinion Mining based
Book Recommender
Systems
Feature based opinion
mining technique 5 [1], [37], [36],
6.
Evaluation Strategy for
Recommender Systems
User feedback based
evaluation of
recommender system
6 [38], [39], [40]
11 | P a g e
Chapter 2
An Overview of Recommender Systems
2.1 Introduction:
The recommender systems (RS) have grown exponentially in recent few years and its
applications have spread over various domain of life including online shopping of
books, home appliances, movies, electronic gadgets, recommendation of doctors and
hospitals for patients, institute recommendation for students and teachers, hotel
recommendations for tourists and so forth. The philosophy behind the success of
recommendation technology is the fact that it is human tendency to rely upon
experiences of their neighbors and friends prior to making decision of any kind,
especially regarding purchase of any items, taking admissions in institutes for higher
education, opting an apartment for rent or buying it, spending weekend at some
holiday places, etc.
The advancement of Internet technologies has caused data overload due to which
the buyers face more difficulties in finding the exact destination which meet their
needs out of a huge collection of the available options. If a student who wishes to
spend his/her vacations at some hill stations and would like to stay in a hotel with
peace and calm, there would be thousands of places all around the world which might
come to him/her as options. In such a situation recommender systems can provide a
better option according to the need and requirement of the user and depending upon
his/her prior preferences.
Although there are several definitions which researchers have suggested for
recommender systems, we define recommender systems as –
“Recommender systems try to identify the need and preferences of users, filter the
huge collection of data accordingly and present the best suited option before the users
by using some well-defined mechanism.”
In this chapter, we have reviewed more than 100 articles related to recommender
system including the manuscript in which very first existence of collaborative filtering
has reported in mid 90s [41], [42].
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
12 | P a g e
2.2 Previous Review Studies
The first paper on collaborative filtering (CF) was introduced in mid of 90s [42], [43].
The proposed CF technique provided a platform to design recommender system and
laid a strong foundation for the development of such recommender systems. The work
in the concerned area has been reviewed extensively in the literature. The study of the
surveys and reviews of recommender systems helps in establishing a better
understanding of the subject and gives a holistic picture of the technology used in the
field along with various aspects related to the topic. In this section, we have tried to
include major review/survey papers on the related work and discussed their
contributions. As the origin of the recommendation techniques are in mid 1990s, it
seems adequate to include papers from 2000 onwards.
In 2000, B. Sarwar et al. [44] has analyzed the effectiveness of recommender
systems on actual customer data from an e–commerce site and compared several
recommender algorithms with respect to their performance [45]. In 2001, Schafer et
al. [46] have examined traditional marketing methods and provided a foundation for
the growth of recommender systems as a marketing tool for e-commerce. They have
also presented taxonomy for recommender system and identified five models of
recommender applications. One of the excellent contribution of the Schafer et al. was
their exploration of four different domain for future study based on the taxonomy that
have not been adequately explored by the existing applications, then. They have
suggested following four area of research for recommender systems; non-
personalized, attribute based, item-to-item correlations, and people-to-people
correlations.
In 2002, R. Burke [47] investigated possible extent of hybrid recommender
systems and provided quantitative results for relative comparisons. Burke [48] has
also contributed for the researchers by surveying the hybrid recommender systems.
He has made comparison between different recommendation techniques and
hybridization strategies. Four techniques for recommendation and seven strategies of
hybridization were considered. He also has included 41 hybrids with some new
combination of that time. The attraction of the researchers towards recommender
system has been noticed increasing rapidly in early decay of the millennium. The
generations of recommendation by early decay of the millennium has been reported in
[42]. The authors have also presented an overview of recommender systems with the
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
13 | P a g e
discussion of the limitations, and possible enhancement for the solution of existing
issues.
In 2007, Candillier et al.[49]has reviewed the primary collaborative filtering based
systems and done an extensive comparison using MovieLens data set. Their study
identifies advantages and drawbacks of the approaches under evaluation. However,
there was no much discussion about the various issues encountered in the
collaborative filtering based approaches. The issues like data sparsity, shilling attacks,
synonymy, scalability, etc. are discussed comprehensively by X su and
Khoshguftaar[50]. They have proposed possible solutions for the existing issues as
well. The authors have also presented a comprehensive survey for collaborative
filtering techniques, categorized collaborative filtering algorithms and analyzed their
predictive performance in addressing these issues. The evaluation of the recommender
system has been discussed in [51]. The authors have discussed the ways to compare
recommenders based on the basis of a set of properties and described how can
recommender systems' performance be compared for the relevant area of application.
They have described experimental background suitable for deciding preferences
between several algorithms. They have also discussed how to draw reliable
conclusions from the conducted experiments.
In 2012, Park et al.[52] and Zhou et al. [53] have done a good work. Park et al.
reviewed 210 research articles related to recommender system and examined the
research trends in the concerned area by observing publication of the paper year-wise
and journal-wise. The effort helps the interested people with insight for future
research direction. Zhou et al. 2012 also presented an overview of state-of-the-art for
developing personalized recommender systems in social networking environment in
the same year. The work provides a research direction to address user profiling and
cold start problems.
The maximum number of research papers for the survey has been included by
Bobadilla in his tremendous work [54]. They have proposed a method which gives a
criterion for the inclusion of research papers of the concerned field. They have
discussed the overview of recommender systems and collaborative filtering methods.
They have also provided original classification of recommender systems,
suggestedarea of future research including bio inspired approaches for recommender
system.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
14 | P a g e
Table 2.1:A glance of the review studies on Recommender Systems
Serial
no. Author & Year Primary Contribution
Citation on
Google Scholar as
on February 2017
1 B. Sarwar et al.
(2000)
Analysis of effectiveness of RS
comparison of recommendation algorithm 2235
2 J. B. Schafer et
al. (2001)
Provided a foundation for the growth of RS as a marketing tool in e-
commerce
Five models of applications and four domain of future work are explored.
1954
3 R. Burke (2002) Investigated possible extent of hybrid recommender systems
Provided quantitative results for relative comparison. 3153
4
G.Adomavicius
and A. Tuzhilin
(2005)
Generation of RS is discussed
Limitations and possible enhancement are mathematically modeled
7777
(Most cited article
on RS)
5 L. Candillier et
al. (2007)
Reviewed the primary collaborative filtering based systems
An extensive comparison using MovieLens data set.
150
6
X Su & T. M.
Khoshguftaar
(2009)
Issues like data sparsity, shilling attacks, synonymy, scalability, etc. are
discussed, and their possible solutions are proposed.
Comprehensive survey for CF techniques is performed, categorized CF
algorithms and analyzed their predictive performance.
1974
7
G. Shani & A.
Gunawardana
(2011)
Compared RS on the basis of characteristics and application both.
Described experimental background suitable for deciding preferences
between several algorithms.
Discussed method of drawing reliable conclusions from the conducted
experiments.
709
8 D. H. Park et al.
(2012)
Reviewed 210 research articles
Examined the research trends in the concerned area year-wise and journal-
wise.
The effort helps the interested people with insight for future research
direction.
284
9 X. Zhou et al.
(2012)
An overview of state-of-the-art for developing personalized recommender
systems in social networking environment.
The work provides a research direction to address user profiling and cold
start problems.
124
10 J. Bobadilla et
al. (2013)
An overview of recommender systems and collaborative filtering methods
are discussed over 253 articles.
Provided original classification of recommender systems
Suggested area of future research including bio inspired approaches for
recommender system.
683 (Most cited
article since 2011,
Elsevier)
11 J Lu et al. (2015)
It systematically examines the reported recommender systems through four
dimensions:recommendation methods, recommender systems software, real-
world application domains and application platforms.
Provides an understanding of developments in recommender system
applications.
102
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
15 | P a g e
However, they have not discussed about the timing-factor in recommendation and
a little was touched about fuzzy approaches in the recommendation. J Lu et al. in
2015 have systematically examines the reported recommender systems through four
dimensions: Provides an understanding of developments in recommender system
applications.
We have tried to include the discussion on the issue of time-constraint for
recommender systems and also have discussed the fuzzy approaches for the
recommendation of items. We have summarized the work in a tabular form and
shown in Table 2.1.
2.3 Types of Recommender Systems
The recommender systems can be categorized on several bases. In the literature, the
categorization of the recommender systems are usually found [42] on the following
bases;
Approaches used
Area of application for which recommendation is made
Data mining techniques applied, etc.
In [42], RS is categorized in 3 different criteria based on approaches, 1) Content-
based recommendations, 2) Collaborative recommendations and 3) Hybrid
recommendations. Bobadilla et al. [54] have suggested four categories on the basis of
filtering algorithms, Content-based filtering, collaborative filtering, hybrid filtering
and demographic filtering. Burke [47] have categorized 5 types of the recommender
systems based on the approaches. The categories are; Collaborative based
recommendations, Content- based recommendations, Demographic based
recommendations, Utility based recommendations and Knowledge based
recommendations.
We have categorized 8 types of recommender systems (RS). These categories
broadly cover the techniques which have been used by the masses or the current
generation researchers are frequently applying it.
1. Collaborative Filtering based recommender systems (C.F)
2. Reclusive methods based recommender systems (R.M)
3. Demographic Filtering based recommender systems (D.F)
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
16 | P a g e
4. Knowledge based recommender systems (K.B)
5. Hybrid Recommender systems (H.R)
6. Context Aware Recommendation System (CARS)
7. Social network based recommender systems
8. Soft Computing techniques based Recommender Systems
2.3.1 Collaborative Filtering based Recommender Systems
It is the most successful and frequently used recommendation technique discussed in
the literature [55], [44], [56] since the appearance of first recommender system in mid
1990s. The collaborative approach makes use of the recommendation from other
customers whose choices are similar to the target customers (i.e. customer for whom
the recommendation is made). The customers with similar choices are termed as
neighbor.
Table 2.2 :Collaborative Approach illustration
Users Items Purchase
User1 Tv1
Tv2
Tv3
Tv4
Tv5
User2 Tv1
Tv2
Tv3
Tv4
Tv5
User3 Tv1
Tv2
Tv3
Tv4
Tv5
User4 Tv1
Tv2
Tv3
Tv4
Tv5 recommended
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
17 | P a g e
Thus, two major tasks are being performed in collaborative filtering; 1) finding the
neighbor of a customer and 2) exploring the preferences of the neighbors of a target
customer or user. The neighbor of a user can be formed by analysing the past
purchasing behavior of the user and calculating the similarity scores between the
choices of these users. Whereas the recommendation of the neighborhood customers
can be obtained either explicitly in terms of rating which are numerical values within
a specified range, or implicitly with some defined measures. Implicit
recommendations also involve customer‟s feedback. The customer‟s feedback can be
their behavior noticed by the user‟s log information or it can be users‟ sentiments
expressed in terms of their reviews.
e.g. to understand better how items are recommended using C.F, we give a basic
assumption which is supported by a diagram presented in Figure 2.1 and illustrated in
Table 2.2.
Figure 2.1: Collaborative Filtering Approach
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
18 | P a g e
Assumption for C.F: if user1 and user2 have similar ratings of item1, item2 … item
„n‟, they must have similar ratings for item „n+1‟ also. In other words, if user1 has
high rating for item1, item2 & item3, and user2 too has high rating for item1 and
item2 then user2 must have high rating for item3 also.
The researchers have defined C.F differently and categorized in different criteria
based on approaches and algorithms used. Adomavicius and Tuzhilin[42] expressed
C.F in terms of a utility function which tries to predict utility of the items based on the
rating given to the item by other customers having similar preferences as the target
user. They have divided C.F algorithm in two categories. 1) Model based and 2)
heuristic based. The same categorization has been reported in [57]. However,
candillier et al. [49] have given three categories of collaborative approaches. a) User
based, b) model-based and c) item-based.
In user based approach, a set of nearest neighbors is associated to each user, and by
using nearest neighbors‟ ratings on that item, user‟s rating is predicted for the item. In
model-based approach, a set of users groups are constructed and ratings of members
of its group are explored. By using these ratings of an item, user‟s rating on an item is
predicted. Usually in this CF technique, models are created for recommendation.
These models are designed to produce accurate prediction on real data. However, in
item-based approaches, a set of nearest neighbors is associated to each item, and by
using rating of users on items‟ nearest neighbors, the rating on an item by users are
predicted.
Researchers have applied these C.F to design RS for various applications such as
recommending music, movie, web pages, articles and products for online shopping,
etc. [58], [59]. Further, there are several techniques within the above three categories
which researchers have worked on. The work can be classified further on the basis of
different methods and algorithms. The respective criteria and related work is
described in the following section.
2.3.1.1 Item based and User based CF techniques
Item based and User based recommendation are usually performed by exploiting –
Association rule mining between preferences of neighbor of users
Rating
Choice of individuals for varied items
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
19 | P a g e
Similarity in the preferences of different users for common items
Tagging
2.3.1.1.1 Association rule mining between preferences of neighbor of users
Association rule mining has been used extensively in collaborative recommendation.
An association rule based recommendation technique was proposed by Sarwar et al.
[44]. The authors have suggested some association rule for exploring the association
between user‟s purchase behaviors towards items and accordingly the items are
recommended to users. The authors in [60] have investigated the possibilities of
inclusion of association rule mining for collaborative filtering based
recommendations. Since collaborative recommender exploit how similar are the
customers' preferences, it is easy to make personalized recommendations.
However, association rule mining algorithms are designed by keeping in mind the
concept of market basket analysis. Such algorithms are not useful for collaborative
recommendation as there are enough rules which these methods need to mine, which
may and may not be fruitful for the user. Also, other criteria of association rule
mining often lead to create huge number of rules or some time very few rules which
have a negative impact on the performance of the system. The authors have designed
a collaborative recommendation technique to mine association rules for this purpose.
The associations between users as well as associations between items, both are
considered. In [61] authors have proposed scalable techniques based on association
rule. The rules are discovered from usage data for personalization of web to users.
Sandvig et al. have presented a collaborative recommendation algorithm based on
association rule mining in 2007 [62]. They have used k-NN algorithm to prevent
profile injection attack. Their results indicate that the proposed methods have shown
significantly improved performance.
2.3.1.1.2Rating based recommendation
Since a general trend in recommendation is to get rating from a user for available
items which in turn, support other users to find better items. This trend of
recommendation is simply termed as rating based collaborative filtering. However,
rating based recommendation is used in model based recommendation as well, which
shall be discussed in its appropriate place (see section 2.3.1.2).
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
20 | P a g e
PolyLens[63], an extended version of MovieLens, is very helpful in group creation
and management. Basically, PolyLens is designed for smaller group to recommend
movie. Several factors have been considered while designing PolyLens, like
generating group recommendation, evolution of group and its formation, and the
nature of the group to which a user belongs. It uses the nearest neighbor methods and
presents the sorted list according to lowest ratings.
RACOFI (Rule Applying Collaborative Filtering) is proposed in [64] which is a
multi-dimensional rating system. The authors implemented RACOFI Music for
assistance of users who usually prefers to listening music on-line. Their
implementation helps in recommending and rating audio. Authors have categorized
five features of music which generally have impact on users. They have made their
system available on-line since August 2003 at [http://racofi.elg.ca].
Rating system is also used in TiVo [65] which uses 100 million ratings. These
ratings are provided by approximately 30,000 users of different TV shows and
movies. The TiVo recommends the different TV programs to viewers.
Since a general trend in recommendation is to get rating from a user for available
items which in turn eventually support other users to find better items. The authors
have presented [66] a database-driven approach which makes use of the ratings in
item-to-item CF technique. The authors have claimed the ease of implementation and
its applicability in vast range.
2.3.1.1.3 Choice based recommendation
In choice based recommendation, items are recommended by using similarity in the
preferences of a single user for different items. Hayes and Cunningham [67]
developed a music application, „smart radio‟ at Trinity College, Dublin in 2001. The
music application is a web-based which allows users to share music programs. The
authors have used collaborative recommendation techniques and applied streaming
audio technology. The controlled distribution of music on web by the operators is
studied in their work and smart radio is designed to personalize the music programs.
The idea of collaborative filtering is introduced to swap the music programs by
observing the similarities between the users‟ choice. The smart radio is currently
working and has the permission from Irish music rights organization (IMRO).
Iman et al. Presented [68] a choice based technique that makes use of CF method
and extract latent knowledge from user ratings, and ask the user to prefer one of the
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
21 | P a g e
two sample items iteratively presented before them. The technique tries to place the
user in the latent factor space, and those items are selected for recommendation which
is near to the user position. The authors showed their results present better
recommendations. Since, the authors have used latent factor as well, this CF
technique can also fall in model based recommendation, if perceived otherwise.
As online radio has become popular, the authors have designed [69] a mechanism
by which playlist in real-time of listening the audio can be tailored according to the
musical tastes of the listener. The authors have used CF techniques to generate a
playlist in real time. The audience usually has listening history of the music before
listening to a particular one. On the basis of history of the listener, playlist is
recommended to the listener. They have also described the details of the
implementation of the technique.
A choice-based interface is studied for preference evocation during the cold start
phase [70]. The interface is compared with an existing rating-based system. The
authors have shown results which indicate that rating-based interface take more effort
whereas choice based system provides more satisfying recommendations.
2.3.1.1.4 Recommendation based on similarity in the users‟ preferences for common items
GroupLens[58] is one of the earliest developed collaborative filtering based system
which provides filtered online news to member of a group. It eases the process of
finding news articles which a user might like from huge amount of available news
articles.
Pazzani in 1998 [71] has discussed how to learn profile of user interests and how it
could help in the recommendation of web pages or news articles. The author has
mentioned the collaborative approaches and their pros and cons in the
recommendation of information sources to users by taking examples of restaurants.
In 2001, G Karypis[72] has suggested an item based personalized information
filtering technology to explore a set of N items. These N items are matched with the
interest of certain users. The authors have presented a method that first determines the
similarities between the various items and then the similarity is used for final
recommendation of items. The author has shown that the experimental evaluation on
five different datasets is 27% better. The standard collaborative filtering techniques
face great challenges in terms of scalability and performance, especially when there is
a lack of explicit user ratings. To improve the scalability of collaborative filtering,
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
22 | P a g e
web usage mining techniques can be used. However, it affects the recommendation
accuracy.
An improved FolkRank by using item based CF method is proposed by Gemmell
et al. [73]. They came up with a conclusion that item-based CF if mixed with
traditional graph-based approach could enhance the performance in FolkRank. Thus,
it is evident from the work that CF, especially item-based collaborative filtering,
could be proved an excellent way to enhance the performance of a recommender
system [74].
2.3.1.1.5 Tagging based recommendation
A recommendation approach based on tagging, „FolkRank‟, was proposed in [75],
[76]. Authors have calculated the distance from the uploaded resource. These
distances serve as a base in exploring the tag recommendations.
Another tagging based recommendation approach is presented by Zheng and Li
[77][78]. The system is based on CF. Their research has highlighted the importance of
tag and time in the process of recommendation. In general, rating matrices are used in
traditional systems based on CF; however, unlike others they used matrices based on
tag and time relations. The similarities are obtained by calculating tag-weight and
time-weight. The similarity index helps in identifying new neighbor which in turn
give the prediction on the basis of recommendation they made.
2.3.1.2 Model based CF techniques
Model based CF techniques as described earlier used to develop models using several
techniques including machine learning, Bayesian classification, ordering, clustering,
latent information utilization, graph model, etc. Goldeberg[41] have presented a
model based personalized book recommendation technique. The authors have applied
association rule mining and BNs for personalized books recommendation. The
association rule mining is used for exploring the association between user‟s
preferences by observing the borrowed books. The BNs are implemented is designing
the personalization of the RS.
However, the rating is also used in the model based recommendation. A User
Rating Profile model (URP) for rating-based collaborative filtering is [79] presented.
The URP is designed to assign one rating to each item for each user. The author
introduced a generative latent variable model. Each user is represented as a mixture of
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
23 | P a g e
activities of the user by generative latent variable model. User‟s actions help in
generating the rating for each item by observing activity of a user towards an item. A
preference pattern is associated with each activity of the user which supports in rating
of the items.
The author [80] analyzed existing methods in 2004 from machine learning
perspective to predict the rating. The author has shown that many existing methods
which were designed to fulfill the task are simply modified machine learning
techniques. The basic operations like dimensionality reduction, clustering,
classification, regression, and density estimation are performed. New prediction
methods are developed by the author. Marlin introduced a new experimental
procedure which has not been used previously.
The Kim et al. have proposed a machine learning technique to extract the
marketing rule for personalized recommendation. They have used tree induction
techniques, which can be incorporated with data mining techniques to match the
customer‟s demographic details. The proposed methodology helps in fetching the
rules for personalization of advertisement to a buyer shopping on the Internet [81].
One of the issues with collaborative filtering technique is that they are not portable
and is successful for an Internet environment with large computers. Miller et al.[82]
presented „PocketLens‟, a promising collaborative system that works on connected
servers with even palmtop and the results are no more less than the other competitive
techniques. PocketLens is based on CF algorithm which finds neighbor by the use of
5 peer to peer architectures. A shopbot is presented [83]. Shopbot is basically a
comparison shopping search engine which is designed in such a way that it can
exploit freebies to consumers without paying any extra fee. The authors have
suggested an item-item similarity method by using CF techniques. They have
considered the additional provision of providing the cost of the product as well as
their benefit from saving point of view to customers for recommendations.
Bayesian networks (BNs) is used [84] as a classifier for CF. Binary-class data have
been major focus for researchers in earlier model to perform CF task, however, the
authors have applied advanced classifier based on BNs. Moreover, they have not
worked on traditionally synthetic binary data; instead they have used real-world
multi-class CF. they have showed by their experimental results that their proposed CF
model has the performance better than the traditional CF algorithm, especially when
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
24 | P a g e
rating data have relatively more missing rates. Also BNs based CF is robust as it does
not degrade with increase of sparseness.
One of the fastest methods to improve the prediction accuracy without affecting
the running time is presented by [85]. Previously, the adopted approaches used to
compute interpolation priorities separately; however, Bell and Koren optimized the
problem in a way that they computed interpolation weights simultaneously for
neighbor. This method can generate a prediction in about 0.2 milliseconds. And is
equivalent efficient for large scale applications. The Netflix dataset is used for
evaluation.
In 2012, Sahoo et al. [86] developed a personalized recommendations to help the
user when their preference might change with time. The authors have argued that
user‟s behavior is not static and changes over time. They have proposed a hidden
Markov model. The model performs personalized recommendations by correctly
interpreting the behavior of a user in selecting the product. The preference of a user is
modeled as a hidden Markov sequence. Authors claim that the proposed model
outperforms the existing algorithms when the data is less sparse and the user
preference is changing.
In 2013, Yue Shi et al. [31] introduced ranking in recommendations. Due to the
rise of collaborative filtering (CF), the need of learning to rank has emerged. For
improving the ranking of the top-N recommendations, the ranking method could
contribute significantly. The authors have presented the key ideas of different
categories of learning to rank approaches, and illustrated how these techniques can be
extended to specific CF methods.
CF techniques follow the philosophy of one to one, i.e. every user is independent
and uses a single account. However, in a case where multiple users share a same
account may trouble the recommendation using CF. if context is available then CARS
could solve the issue. But it needs context to be illustrated and explained [87]. Author
proposed a solution to solve the issue without being aware of the context, by using top
N shared accounts, an item-based top N collaborative filtering recommender system.
The method gives the recommendations according to the binary positive feedback.
The experimental results show that their techniques can tackle the issues regarding
shared accounts of various datasets.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
25 | P a g e
ExcUseMe[88] is the only pure CF based recommender system which tries to
avoid cold start problem without combining content filtering or context details. The
authors have presumed that the arrivals of users for purchase is randomly sequenced
and certainly system takes the decision about the possibilities of new user
participation in the exploration of newly launched items. The users which are possibly
interested in new items are revealed by ExcUseMe gradually. The new items are
modeled according to the user‟s preferences. The provable guarantee for cold start
problem is assured by [68]. The authors have used matrix factorization. The
theoretical prove of the error estimate is also given [88].
2.3.2 Reclusive Methods based Recommender Systems
It is clear from the above discussion that collaborative filtering is based upon finding
similarities between users. It does not need any representation of the objects to be
recommended. Unlike collaborative filtering, reclusive approach exploits the features
of the objects and requires its representation [89]. The reclusive methods are
considered as complementary to collaborative techniques. And it emphasizes on
finding similarities between objects, i.e. items rather than finding the similarities
between users.
Let us consider the example illustrated by using Figure 2.2. There are five different
TVs for which reclusive approach is described for a user. The user has preferred TV1,
TV2 and TV3 either by purchasing or by putting it into cart. TV4 and TV5 are newly
launched items. The features of TV5 are similar to TV1, whereas TV4 has different
representations in terms of its characteristics. Thus reclusive approach which is also
referred as „content based or feature based recommendation‟ would recommend TV5
to user and not TV4.
Reclusive recommendation or content-based Recommendation [90] mainly came
from the concept of information accessing and is a kind of recommendation method
based on comparing users‟ preferences and associating contents between items in
order to provide recommendations to users. This content-based method is also called
Feature-based Recommendations [91] that judges and find out items users are
possibly interested in by analyzing the attributes and characteristics based on User
Profile. The results are then recommended to users. It could even further assign
different weights [92], [41] based on the degree of association between user‟s
preferences and targeted contents in order to better fit users‟ requirements [93].
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
26 | P a g e
Figure 2.2: Reclusive Approach for Recommendation
Like CF techniques, the recommendation approach for Reclusive techniques can
also be categorized in the three types, i.e. 1) Heuristic based, 2) Model-based and 3)
Web mining based. By the use of model based approaches, reclusive method tries to
exploit different machine learning algorithms, classification techniques like Bayesian
networks (BNs), probabilistic approaches, to group the preferences of users based on
the content of the items purchased. Whereas, heuristic approaches uses different data
mining techniques like clustering, decision tree, rule induction, etc. to fetch the
product‟s features and recommend the one which is closest to the preferences of a
user. A category, opinion mining, is explicitly classified as it has been used frequently
in characterizing the items‟ features. Customer‟s reviews and their log information
helps in making a consensus about the features of an item whether it could suit the
users preferences or not?
2.3.2.1 Heuristic based Reclusive Recommendation
The Kim et al. [81] used decision tree to personalize the web advertisement for a user.
The authors have proposed personalized recommendation techniques for the
customers based on their past purchasing behavior. User profile is maintained to
observe the attitude of a user towards similar products. The authors [94] presented a
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
27 | P a g e
technique that combined the feature of classification; user based collaborative filtering
and association rule mining. The classification technique is used to mine the book
with respect to book‟s features. The latter two techniques are used to know the user‟s
requirement for recommending highly rated books. A book recommendation system
based on digital signage system has been proposed by the authors in [95]. The books
are recommended for particulars by identifying age and sex of the users. Here books
recommendation approach is confined and very limited. It cannot be spread to a big
community or universities but only for few magazines for the user aged 19-21 of same
located schools.
In [96], James and Nick have developed a recommender agent for the
recommendation of movie (available at www.filmrecommendations.co.uk). The
proposed approach make predictions based on content that relates the features
accompanying in a movie like, actors, directors, stories, etc. new movies are included
for making recommendation to users. The authors have improved the accuracy by
their pure reclusive approach.
Kazai et al. have presented a mobile app which is enough intelligent to learn the
user‟s interest from the past purchase history or activity knowledge of user at social
network sites [97]. The app provides users with crowd curated content. The app is
also capable of providing users with the knowledge of contents like by the user of
twitter followed by them.
2.3.2.2 Model based Reclusive Recommendation
The researchers have usually utilized users profile to model it for storing their
preferences. These preferences are matched with the feature or contents of the items.
If there is a match between user interests and product‟s features, the item is
recommended to the user. K. Lang [98] has sorted the user dependence problem in
profiling user‟s preferences. Lang has proposed „Newsweeder‟, a technique that has
the provision for users to rate the news they have read in 1-5 rating scale. This helps
in recommendation of next news for the user. Pazzani[71] has proposed a model
based reclusive approach for fetching the user profile about their purchase. The
authors have also suggested CF and demographic technique and combination of trio
for a better recommendation.
Books, journals and research papers recommendations have helped a lot the people
to fulfill their need and get benefited of the recommendation for their study of course.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
28 | P a g e
Mooney [99] proposed a content based book recommendation technique, called
LIBRA (Learning Intelligent Book recommendation Agent) that utilized information
extraction and a machine learning algorithm to explore the features of the books in
recommendation. Jomsri[100] proposes a library book recommendation system based
on user profile loaning and association rule. This system is useful for particular
resides in the same institute within the same library and campus. The experiment is
performed for the specific university only.
A user interface is designed for wireless information devices [101] by using user
feedback. User interests learning model are framed for the current events through
news. Machine learning methodology based on reclusive approach is developed. The
authors have claimed their system can adapt according to the interests shown by the
users. Also, the information size is reduced by the methods; as a consequence, users
can save for obtaining the relevant information.
Since, reclusive approach tends to recommend those items which user has already
aware of. This leads to the problem of overspecialization. The authors [102] have
presented mechanism that overcomes overspecialization. Firstly, by exploring
knowledge of user‟s preferences, then matching the preferences with launched items
at shopping sites.
Bansal has proposed content driven user profiling [103]. The system provides
recommendations for news and blog articles. The recommendation is supported by
Comment-valued approach using topic modeling. A novel hierarchical Bayesian
modeling approach is combined with classical recommendation technique. The
content based solution also exploits user profiles which are enough influential in
providing personalized ranking for users of comment-worthy articles. The system
handles with cold-start issue with no extra requirement of meta-data.
2.3.2.3 Web Mining based Recommendation
As discussed in Chapter 1, web mining techniques are sensibly useful in processing
the web data for extracting the desired information and performing operations
according to the need of the problems. Web usage mining, web content mining and
link mining i.e. web structure mining; all three leading web mining techniques are
used in recommender technology recently. Since reclusive approach, mostly referred
as content based approach, exploits user profiles and items descriptions to guess what
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
29 | P a g e
user could like in future, depending upon the past preferences of a user, irrespective of
the choices made by other users. Most content-based recommender systems encounter
those ambiguities which usually a natural language suffers. The authors [104] have
presented comprehensive methodology to overcome the issues which is associated
with keywords based approaches.
Cho et al. [61] proposed a personalized recommendation system which is based on
Web usage mining. They have suggested an improved collaborative recommendation
methodology which can enhance the quality of recommendation for an Internet
shopping mall. Further, sparsity and scalability are addressed well here to overcome
the poor recommendation problems. Another personalized recommendation based on
Web usage mining is proposed by Kim et al. [105]. Their method is mainly targeted
the problem of helping customers to achieve recommendation only about the products
they wish to purchase. Kim et al. have experimentally evaluated the proposed
methods by applying it on a shopping mall of Korea.
A detailed discussion about the development of a personalized product
recommendation system based on customer‟s click streams is performed in [106]. The
authors have proposed a recommender system based on web mining to overcome the
problem of data overload so that satisfactory recommendation can be made for users.
Web mining techniques are used to observe the purchase behavior of the users and
adopt the change in the users‟ preferences dynamically.
Although there have been good number of studies on opinion mining, however few
of them lead to products recommendation. User feed-back based recommendation for
electronics items are performed by the authors in [39][38], [40]. Liu et al. [107]
proposed a novel product recommendation methodology by combining group decision
making and data mining techniques. It addresses the customer lifetime value (CLV) to
a firm. The authors in [1] recommended books for online shopping using web
mining technique where they categorized the features from the reviews of the users
available online and recommended top computer science books by assigning weights
to these features and scoring these values. The authors in the paper searched the book
on a specific topic using Google search. The top links are stored and the reviews of
the readers for all the stored results are assessed. The features are extracted from the
user‟s review and accordingly the books are ranked.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
30 | P a g e
The reclusive methods are very effective in recommending TV program [78] as the
content of a TV program can easily be traced by the features of programs like time of
program being telecasted and characters involved in the programs, etc. The reclusive
approaches can be a solution to sparsity and cold start problems to an extent. Authors
have suggested reclusive approaches in music recommendation to overcome these
issues. In [108] reclusive approaches are proposed to overcome the sparsity while
authors [96] used reclusive methods to solve the cold start issues. There is few music
systems developed to recommend music to a particular group [15].
2.3.3 Demographic Filtering based Recommender Systems
The recommender systems based on demographic filtering also use similarity
measures as a metric. But instead of finding similar rated items by neighbor users, it
tries to find the similarity between users‟ demographic information like, age, sex,
occupation etc. In this approach, the system stores the demographic information of the
customers and whenever a new user comes to merchandisers‟ site for the purchase of
any product, the system identifies the similarity between user‟s demographic
information. According to the preference of the customer, the system recommends
alike items to new user having similar age, sex, occupation etc. to customer. A typical
recommendation approach of demographic filtering based recommender system is
shown in Figure 2.3
In the figure, four different users are shown, user 1 and user2 are from same region,
they both are teenager students from France, i.e. both the users have almost same
demographic values. Whereas user3 and user4 are from different region with different
occupation and they belong to different age group. However, both are females. Thus
user 3 and user 4 differ significantly from each other as well as from user 1 and user
2. Hence, once the purchase and demographic record of user 1 is stored, the system
would likely to recommend same item to a new user (say user 2) who is common in
various or all aspects with user 1. Also, it is important to decide what types of
similarities between the users are desirable. As we have seen above there is significant
difference between user 3 and user 4. However, both are female and hence they may
have similar choices of buying a product (like clothes, food products, etc.).
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
31 | P a g e
Figure 2.3: Demographic Filtering based Recommendation Approach
Thus, the choice of one of the users can be recommended to another on the basis of
partial demographic similarities. Demographic information can be useful in finding
the category of users whose choices are similar for certain objects. Krulwich et al.
proposed LifeStyleFinder [109] and have used 62 clusters of users which were pre-
existed and has made recommendations to users on the basis of other users belong to
the defined clusters. Pazzani [71] attempted to apply minimal effort for collecting
information of users, and classified users using text classifications. They have used
hybrid techniques including collaborative, content and demographic information for
making recommendation of Restaurants. They have come out with a conclusion that
demographic methods can help in finding evenness in the descriptions of users that
have similar choices of the restaurants. Content-based methods find evenness among
the details of restaurants preferred by a particular user. However, collaborative
method helps in finding correlation between the user‟s ratings of a particular
restaurant and the user‟s ratings of other restaurants. Their experiment demonstrated
that the consensus-based method is effective than any one of the individual method
discussed above. Usually demographic filtering based recommendation technique
when hybridized with collaborative or reclusive recommender approach is found to be
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
32 | P a g e
more effective. We call it as hybrid techniques. The hybrid technique is discussed
later in the section. These combinations can work well and may solve cold start
problems to an extent.
Laila et al. [110] presented a solution to cold start problems which occur while
using rating history of the user as a base for recommendation. They have used user‟s
demographic details and combined it with reclusive and collaborative approach to
provide recommendation for new users with no prior preference and rating details. A
significant impact of demographic information of users for recommending research
papers have been reported in [10]. Kim et al. [81] have suggested demographic
filtering based recommender system. The filtering is based on decision tree induction
and machine learning techniques.
2.3.4 Knowledge based Recommender Systems
The recommender system has much of its emergence due to the initial involvement of
collaborative filtering methods. However, later a good amount of work is contributed
using reclusive methods too. The early implication of collaborative and reclusive
approaches to the recommendation technology has given a distinguished identity to
the above two techniques in the categorization of recommender systems. As
recommender system is a knowledge based approach, thus all the different categories
are based on knowledge filtering techniques. The reason behind keeping reclusive
and collaborative as a separate category is its familiarity and domination from early
days of evolution of recommendation technology.
Apart from the above two techniques, collaborative and reclusive, any
recommender technique by default may be inferred as a knowledge based approach.
However, demographic filtering is based on collaboration of users‟ demographic
knowledge; it seems adequate to keep it as a different criterion. The idea which
differentiates knowledge based systems from other systems, is the degree of
importance it gives to the following two domains –
a) user‟s requirement
b) Characteristic of the recommended items.
The above area of expertise helps in achieving users‟ satisfaction by fulfilling their
needs. Certainly, an approach for building recommender system which needs either
explicitly defined set of recommendation rules or some sort of similarity measures
from prior purchase history of the users is perceived as knowledge based approach for
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
33 | P a g e
recommender system. It is important to know the knowledge sources while
categorizing a recommender system. It is more difficult to concisely typify knowledge
based system than collaborative or reclusive systems. However, the recommender
systems that use supplementary knowledge sources which are not exploited by
collaborative and reclusive recommendations can be characterized as, “knowledge
based systems”. These systems depend more upon knowledge sources, while others
frequently-used techniques do not depend highly upon such sources of knowledge.
Towle and Quinn [111] have argued that an additional information provided by the
user could help in overcoming the sparse related problem as well as cold start
problems. Hence, instead of „rating‟ based recommendation, which is an implicit
approach, they have suggested explicit model for recommendations. The authors have
configured three major retardant in the sensible success of recommender systems.
First, customers show reluctance for receiving recommendation if there is not up to
the mark recommendation constantly, second, the constant arrival of new items and
third, all the products do not have same characteristics. Thus, explicitly asking the
requirement and choices from the user would allow training the system according to
the user‟s need.
Knowledge based approach is applied in [112] for recommending programs on TV
to a group. Since most of the RS need explicit ranking from the users. Merging these
individuals ranking to one consensus ranking so that it may suit all member of a group
well is a tough job. The authors proposed a method which learns the family
preferences separately. On one hand the method keeps the privacy of the family
preferences and on other hand it adapts to the changed preferences of a family. The
classifier is applied to adapt the preferences of each family separately. A recall of
0.57 and precision of 0.30 have been achieved by the author‟s suggested work,
although not much description of the Meta data was provided.
A similar approach is presented by Yu et al. in 2006 [18]. The authors have
provided recommendation mechanism for TV programs for a group by exploiting user
profile. The selected strategy first merges all user profiles to construct a common user
profile, and then uses a recommendation approach to generate a common program
recommendation list for the group according to the merged user profile. The total
distance minimization is used for evaluation of the results. The system works well for
group of users viewing the TV together.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
34 | P a g e
For tourism recommendation; aspects like the charming of a place, ease of
accessibility and accommodation, and well-furnished restaurants are often seen as
important factors. Entrée, a FindMe driven system proposed by Burke et al. [113],
[114] to recommend restaurants by using knowledge based approaches. The authors
have clubbed concepts of several retrieval strategies involving knowledge based to
fetch the destined information. RentMe system is designed which follow the
guidelines of FindMe system for the recommendation of apartments in Chicago on
rent.
To explore the best suited locations of restaurants for a group of people, a
recommender system, „Pocket-Restaurant-finder‟ is suggested in [115]. It
incorporates the choices of the associates of a group. Furthermore, the application
developed can help the group members in real life and have been designed to run at
any kiosk to help a group in finding the restaurant of their mood. A system,
Collaborative Advisory Travel System (CATS), has been presented as a solution for
the recommendation of holidays. The system also tells the area where these holidays
can be engaged.
SPETA [21] is a recommender system behaves like a guide that provides the
service to tourist by observing their past preferences and locations. The suggested
system makes use of the knowledge of user associated information like current and
past locations and preferences. The information for the user is extracted which is
integrated with innovative techniques to provide pleasant experiences to tourists. E-
learning courses have been recommended with different techniques, proposed by
authors in [12], [13], [116]–[118]. In these works authors have proposed
recommendation methodology for courses to graduate students at university level and
for online learning environment. A course recommendation for open university of
China is proposed in [117]. In [119] authors have used machine learning technique to
recommend courses for new enrolled students.
Further, two different types of knowledge based systems are reported in literature
[120], [121], [122], [114], [123].
Case based recommender system.
Constraint based recommender system
These recommender systems are described below.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
35 | P a g e
2.3.4.1. Case-based Recommendation:
In case-based approach, recommendation is largely perceived as a problem of
evaluating resemblance of a product with user‟s preferences. The approach employed
in Case-based recommendation is somewhat similar to reclusive approach in an
exceedingly sense that both the approaches need detail descriptions of the products‟
features. In turn, these features are matched with the user‟s preferences to best suites
their requirement and provide a high level of user satisfaction. Since the requirements
and preferences of users aren't well outlined, hence, similarity-assessment method
helps in up the standard of the recommendations, this is why case-based approach has
gained a great success in e-commerce [124].
Let us consider an example [125]. If I go to market for buying refrigerator, the
seller may and may not be acquainted with my preferences depending upon whether I
have made my purchase from there before or not? Obviously, if I have purchased
refrigerator before, why should I go again? That is seller is not aware of my
preferences, right? Now, if there is description of the products like company to which
it belongs, size of the refrigerator, color, power consumption and warranty durations,
it will help the seller in providing the closest to choice object for customer. Let I was
provided an item whose similitude to my preferences are high but I dislike the color.
“Everything is fine but may please you show me a blue of it?” it would be my request
from the shopkeeper to give me an item with similar features but the color should be
blue, with this additional explicit knowledge provided an exact recommendation can
be made with less effort and time. This is what a case based recommendation does in
recommending the items to users. Case-based recommendation treats
recommendation as primarily a similarity-assessment problem. How can the system
find a product that is most similar to what the user has in mind, with the
understanding that what counts as similar will often involve domain-specific
knowledge and considerations?
In product recommendation, decision trees have been used extensively. McSherry
[126] has come up with an idea of treating the decision tree as an identification
method which identifies an item as an object for recommendation and stores it in case
library as a single case. The authors have tried to reduce the complexity in acquiring
the explicit knowledge from user for case based recommendations. McSherry in
another work [127] has talked about how recommender system is affected by
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
36 | P a g e
incremental query elicitation. Generally, obtaining the additional knowledge from the
users, hinder in obtaining quality solution. The context for which the dialogues can be
stopped without any loss in quality of solution is explained by the authors. It is
suggested by the authors that destination-oriented technique in which number of cases
gets dominance over target case, would provide a better solution. Further, it is noticed
that the strategy costs less in computation as well. The authors have evaluated their
results on Travel case library (TCL). TCL is a standard benchmark which contains
more than 1,000 cases. It is found that their method reduces the average length of
argument better than others.
The explanation of recommendation that briefs the user why recommendations
have been made would attract the users and might satisfy for a good extent [128].
With this principle in mind, authors have not tried to justify the specific suggestion
but rather explained the reason of suggestion. The philosophy also helps users in
knowing the further opportunities in a case when the recommended items dissatisfy
them. The compound critiques are trained to work as a form which may generate
feedback. The authors have claimed explanation-rich critiques improve
recommendations for users.
2.3.4.2. Constraint based Recommendation:
To understand constraint based recommendation, let us consider an example of how
recommendations are made for web hosting services [121]. The personal preferences
regarding cost, bandwidth, visitors count, etc. are required to be provided with users.
The recommender suggests the users and explains the reason of recommendation on
the basis of the preferences of the users observed. If no solution can be acquired by
the recommender, a replacement is required to be provided for users, in order to save
the users from going into a dead end situation. The above example is a better
explanation for a constraint based recommendation [129]. In these recommender
systems, features of the product and association of user‟s requirement with these
features, both are modeled in the form of a constraint. Constraint-based approaches
help in purchasing the items which are not frequently purchased. Constraint-based
recommenders support customers in a deadly scenario where no other solution is
provided by automatically suggesting options for remedies and explaining
technicalities with the items‟ features.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
37 | P a g e
The application of constraint-based recommenders for financial services is
presented in [130]. Another financial application of constraint based is reported in
[131]. The authors [132] present an approach to enhance the recommendation for
multimedia. The additional feature for component visualization is associated with
constraint based. It enables users to interact the virtual product directly. Visualization
functionalities provide substantial contributions to user-friendly interfaces boosting
the acceptance of recommenders.
Knowledge-based recommender technologies [130] enable customers and sales
executives to identify the appropriate products and services. These knowledge
engineering are also useful for complex and high involvement products such as cars,
computers, or financial services. The authors have presented the VITA
(VirtualisTanacsado) financial services recommendation environment which has been
deployed for the Fundamenta building and loan association in Hungary.
The effective integration of configuration system development with industrial
software development is crucial for a successful implementation of a mass
customization strategy. On the one hand, configuration knowledge bases must be easy
to develop and maintain due to continuously changing product assortments. On the
other hand, flexible integrations into existing enterprise applications, e-marketplaces
and different facets of supply chain settings must be supported. The authors have
designed a model-driven architecture (MDA) for model development and interchange,
and sensibly argued how the industrial configuration can serve as a foundation for
standardized configuration knowledge representation; thus providing knowledge
sharing in heterogeneous environments. [131].
The problems with DVR and catch-up TV has been resolved by methods proposed
by [133]. The challenges and solution regarding personalizing the topic have been
illustratively explained in this work. The author has concluded that there are the
contents which are absorbed sequentially trends of seasonal dynamics is observed
with these contents. If new content arrives just after broadcasting of any content, it
would lead dynamic stream of data. And there may be repetition of similar data for
different services simultaneously.
2.3.5 Hybrid Recommender Systems
Though Collaborative Filtering (C.F) and (R.M) are the most frequently used
techniques in designing Recommender Systems (RS) but they inadequately provide
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
38 | P a g e
any explanation of why the specific recommendations have been made to particular
user along with recommendation, hence, they fail in fulfilling the explanation in
various scenarios. These shortcomings of the both leading technologies can be
overcome by the use of the combination of duo. The various combinations of these
techniques have been presented in the literature. These combinations are termed as
„hybrid technique‟. We have categorized seven types of hybrid recommender systems
based on different combinations –
i) Hybrid Recommender Systems based on Collaborative Filtering (CF)
dominated Reclusive Method (RM)
ii) Hybrid Recommender Systems based on RM dominated CF techniques
iii) Hybrid Recommender Systems based onunified RM and CF techniques
iv) Hybrid Recommender Systems based onSubsequent Integration of
separately applied CFtechniques and RM
v) Hybrid Recommender Systems based onIntegration of CF and RM with
knowledge based system (KBS)
vi) Other Hybrid Recommender Systems using CF techniques
vii) Other Hybrid Recommender Systems using RM
The work which incorporates these combinations has a wide range and has been
applied over various applications. The techniques of the hybridization are described in
the following section.
2.3.5.1. Hybrid Recommender Systems based on Collaborative Filtering dominated
Reclusive Method
Incorporating components from CF and RM lead to form a hybrid recommender
system. These hybrid recommender systems help in dealing with the above said
shortcomings. The researchers have started to explore the frequent occurring
problems with these two techniques, namely overspecialization and cold start
problems. The hybrid technique, composed of the combination of these techniques in
various suitable combinations is proposed, and a new approach for recommender
system is perceived. It is noticed that various aspects which should be retained in the
designing of recommender systems, are ignored. Not considering these aspects may
dissatisfy the users and the ultimate goal of the recommender system cannot be
achieved. The advantage of employing hybrid system is the power of assimilation of
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
39 | P a g e
these methods in integrating the collaborative and reclusive approaches by
contemplating the best of the two without considering the drawbacks of the either
[134].
Mooney et al. have [99] presented an effective methodology combining content
and collaboration. The content is used in enhancing user data whereas personalization
of recommendation is made through collaborative filtering. The hybrid system
performs better than pure CF technique or Reclusive Method [108].
Content with collaboration are elegantly combined in [108]. The authors used
reclusive approach to design feature-based predictor for boosting the user profile. CF
techniques have been utilized further to provide personalized recommendations. The
authors have shown that Collaborative Filtering dominated reclusive methods
outperforms the pure reclusive recommendation, pure collaborative filtering
techniques, and simple hybrid approach.
The Good et al. have come up with a conclusion that CF techniques can be
combined with content based agents, which in turn gives the best recommendations
than any combination or separate techniques would produce. They designed the
system in such a way that users need not to choose best in agents, instead, the CF
framework recommends best ones for them [45].
A clustering technique has been presented [135] as a solution to cold start
problems. Item-based CF techniques make use of clustering strategies. The
comprehensive idea of integrating the content information into the CF has been
explained. The authors have used MovieLens data for experiments. The results
evinced the improvement for cold start problem.
The method to combine features of human personality into the traditional rating-
based approach for CF systems is presented by Hu and Pu [136]. The rating based
system usually computes similarity of the users‟ preferences with its neighbor and a
naïve user may not find good recommendation due to lack of exploration about their
past preferences. Combining human characteristics with CF techniques provides
better recommendation for new users whose past rating preferences are not well
formed.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
40 | P a g e
2.3.5.2. Hybrid Recommender Systems based on Reclusive Method dominated Collaborative
Filtering Techniques
RM dominated CF techniques implies those hybrid systems which incorporate CF
techniques into Reclusive approaches. The basic philosophy of reclusive approaches
is retained and collaborative techniques are applied over there. A technique for the
purpose of text filtering by combining collaborative and content methods are
presented in [137]. The latent semantic technique is used for storing user profiles. The
RM dominated CF techniques performs well than the simple reclusive approaches.
Since, collaborative filtering methods are treated as a base in recommendation
technology. It utilizes the recommendations based on other users‟ preferences. By
contrast, reclusive approach is powerful enough to make recommendations by
obtaining details about an item. Thus, reclusive approach can recommend items which
are not previously rated by user. The additional feature of CF techniques for getting
user profile stronger can boost the recommendation process if the two techniques are
combined [108] . The authors have presented the results which demonstrate that RM
dominated CF techniques can give correct recommendations [99].
A combination of RM and CF [138] is used to recommend TV program to viewers
of Ireland and Britain by collecting their rating and reviews. The authors have
discussed a content personalization system which selects the most suitable contents
from an individual by reclusive dominated collaborative approach. The key to
address the issue is the exposure of a learned user profiles. The duo combination
provides a vigorous personalization solution.
2.3.5.3. Hybrid Recommender Systems based on unified Reclusive Method and
Collaborative FilteringTechniques
Several authors have integrated CF and RM in many ways [54] . However, coalescing
the two methods into one, is proposed by Ansari et al. [139]. As the brand online
retailers like Amazon, eBay and Yahoo! use CF or Reclusive methods for the process
of recommendation of various products and services to their users, unifying the duo
could enhance the recommender strategies [66]. The authors have described a
Bayesian model to sort out the preferences by categorizing the information into five
different types of knowledge associated with the characteristics of the recommender
system concerned. Markov chain Monte Carlo methods used to recommend movies.
The Monte Carlo model works in all circumstances whether CF technique are able to
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
41 | P a g e
be employed or not. Thus, in general, a recommender system can predict whether a
particular product or service may fall into the preference category of a user, in
addition it can guess for a user the movie he would be interested in, for sure. The
inductive learning method [141] is proposed which incorporates the characteristics of
the artifact, which is utilized by the recommender systems in making predictions.
A novel approach for the recommendation of on-line academic research papers
based on ontology to help in boosting the user profiling is discussed [142]. The
authors collect feedback for users‟ profiles by utilizing a novel approach based on
profile visualization. The ontology of the research papers topic support in categorizing
the papers which in turn serve as a base for collaborative recommendation. The users
who have similar preferences in browsing the research papers of same interest are
stored and accordingly the recommendations are made.
A unified framework for collaborative and reclusive recommendations based on
probabilistic method is discussed [143] which is an extension of Hofmann‟s aspect
model [144]. The method assimilates item‟s content with users and items which is
generated by data source itself, and provides a solution in recommendation when data
are prevailed by sparsity.
2.3.5.4. Hybrid Recommender Systems based on Subsequent Integration of separately
applied Collaborative Filtering Techniques and Reclusive Methods
The separate implementation of RM and CF are applied in [137] and [145] where the
authors have discussed the improvement in quality of recommendation. Kim et al.
[146] have proposed a book recommender system for the validation of their method
which was designed for an online community. They tried to satisfy the minor
members of group which are left unsatisfied although the majority may have
satisfaction due to the differences in preferences.
The recommendation technologies have been useful for recommending courses as
well as utilizing the courses for other library management program. In [147] authors
have used academic courses to generate data for library planning purposes.
Billsus and Pazzani have proposed the induction of hybrid user models. The hybrid
model comprised of separate models for RM and CF techniques. The detail
description of the implementation of these algorithms for addressing the issues
booming in recommendation technology is done [148]. The CF and RM are also
integrated and discussed from the restaurants recommendation perspectives and
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
42 | P a g e
illustrated the advantages of the combination over separate implementation of either
of the techniques [71].
2.3.5.5. Hybrid Recommender Systems based on Integration of Collaborative Filtering and
Reclusive Methods withKnowledge based System
The use of combination of CF techniques and RM with knowledge based systems
(KBS) has been reported in literature. The authors [149] have suggested a
recommendation technique which identifies sets of rule and deduce the
recommendation upon these rules. This recommender system provides accurate and
cheap clinical examination to patients. A recommender system which provides
personal health information of users is designed by Lee at al. [26]. It uses profile of
users and accordingly the information is provided for the better services of patients.
Wiesner et al. [150] kept the fact that physical activities are very important for fitness
and health, so, they have designed a physical activity recommender system that tells
the exercise time useful for people. Recommendation of nutritious diets have been
suggested in [25]. They have used user ratings for the nutrition needed accordingly
provided the nutritious diets to users. The usage of RS in health has been explored by
Fernandez et al. in [151], where a detail of RS and their extensive uses in the domain
of Health and care is discussed.
A recommender system has been suggested in [84] for e-business by introducing
computational ecologies. This system supports recommendation based on negotiation
which also inspires ecosystems monitor [152] .
For private banking, a recommender system „PB-ADVISOR‟ for multi investment
has been framed. The system addresses the issues of recommendation with
explanation, in addition it also generates several packages and has the ability to
suggest best services for customers with appropriate explanations [153].
2.3.5.6. Other Hybrid Recommender Systems using Collaborative Filtering Techniques
Knowledge-based (KB) and collaborative-filtering (CF) recommender systems, both
have equally contributed online recommendation for users to find products close to
their choices out of a huge data with large varieties. R Burke in 1999 has explicitly
described in detail the pros and cons of these two [123]. The author has outlined the
chances of collaborative and knowledge based hybrid recommender system. In the
suggested methodology knowledge-based techniques and CF techniques both work as
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
43 | P a g e
a complementary for others. KB technique bootstraps the CF engine, and the CF
filters the KB recommendations.
A film recommender agent expands and fine-tunes collaborative-filtering results
according to filtered content elements - namely, actors, directors, and genres. This
approach supports recommendations for newly released, previously unrated titles.
Directing users to relevant content is increasingly important in today's society with its
ever-growing information mass [96]. Tang et al. [154] have suggested QoS services
by using a hybrid techniques which combines CF method with location aware
approach.
The use of CF techniques with temporal dynamics [155] is studied. The authors
have presented a hybrid recommender system comprised of CF techniques and graph
based model [156]. The graph-based approach has already been proven superior to
other methods by experimental results. In other words, what information has been
conveyed by other techniques is already suggested graph based model. An extensive
evaluation has been performed by authors.
2.3.5.7. Other Hybrid Recommender Systems using Reclusive Method
The CF and RM are the base technologies in recommendation. The most of the
technology either use both techniques in any combination or combine either of the
technique with other methods like knowledge based approaches, context aware
approaches, etc. since contents are used in reclusive approach to extract the features
associated with the items in consideration, it would be a powerful combination if
reclusive methods are combined with knowledge based approaches. A personalized
recommendation could be a solution in providing user the matching items to their
preferences out of the huge data available. A hybrid method which combines the
reclusive approach with knowledge-based methods to enhance the recommendation
performance is presented [157]. Explicit and implicit feedbacks are taken from the
users for recommendation process. Optimized weight vectors and preference matrix
(PM) are used for exploiting implicit and explicit attributes respectively. The hybrid
system gives better results in reducing cold-start and sparsity.
The reclusive approach is able to recommend users the product which have been
already searched or visited by them and cannot predict about one which has no past
record. However, in many cases users may wish to go for purchasing a new item they
never have seen before, as the unheard items may be of interest for a user. The
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
44 | P a g e
situation is termed as serendipity. Incorporating serendipitous recommendation
strategy with reclusive methods alleviate the over specialization problems in
recommendation [158]. The authors have suggested hybrid recommendation approach
to recommend surprisingly the new items to users. The hybrid methodology is
comprised of reclusive and serendipitous approaches.
In [23] the authors have tried to use sematic web structure and text mining
techniques for providing users the risk that may occur if the ignorance are kept alive.
Thus these risks are advertised on social networks, etc. A hybrid music
recommendation system which handles the issues encountered with collaborative and
reclusive approaches has been reported [159]. The authors have utilized the rating as
well as content of data by using a Bayesian network. The approach solved the
problems of collaborative approach of not being capable of recommending music for
which no ratings have been recorded. In addition, it also resolves the issue in studying
the artist varieties. Latent variables are used to explore the solutions [16].
2.3.6 Context Aware Recommender Systems
Context aware recommender system though can be perceived as a special kind of
knowledge based system, when context is involved as knowledge, required for
recommendation. However, the high inclination of the recommender system research
community towards recommender system for learning has provided a platform that
compels us to keep CARS as a different category, and not a type of KBS.
The ultimate goal of recommender system is to achieve user satisfaction. And user
can only be assured for their satisfaction if they are delivered with the exact
recommendations that meet their needs. The user‟s requirement is not static and may
vary time to time depending upon various social and other factors affecting their
purchasing trend. Hence, Figure 2.4shows a context aware recommender system
(CARS) which takes into account the context in which user goes for some specific
item. And different varieties of the items can affect the user‟s demand significantly.
We illustrate the example with the help of Figure 2.5. Let us consider that user
needs to buy clothes from a cloth merchandiser, obviously the demand of cloth for the
type it belongs must depend upon the season and weather. In winter season user must
be asking for the woolen cloth. Now, if we consider the reclusive approach it would
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
45 | P a g e
Figure 2.4: Context Aware Recommender Systems Overview
Figure 2.5: Example for Context Aware Recommender Systems using season based clothes
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
46 | P a g e
be recommending woolen-like clothes to the specified user always, irrespective of the
context. In collaborative approach, the system would go to observe the user‟s
neighbors preference, eventually in the scenario, the probable recommendation may
be clothes similar to woolen.
The context aware recommendation is necessary to understand the user‟s delicate
preferences and exploiting the complications in their requirement explicitly. It is
shown in the Figure 2.5 that how a CARS would care for user‟s choices. A Context
Aware system would explore the situation in which the user‟s purchase is noticed, and
tries to filter the recommendation accordingly. Thus, in a summer season, CARS can
never recommend a woolen cloth to user. However, in the same scenario, the other
systems may exhibit false positive error. False positive employs the recommendation
of an item to user while the item is not needed to be recommended and not preferred
by the users.
In mobile environments there can be various contexts needed to be considered
while making any recommendations. The considerable context can be weather, time,
route, location, ad transportation means, etc. before making any recommendation in
such scenario recommendations should be designed context-aware for guiding the
users on mobile path. A context based travel-related information for mobile systems
are proposed [160]. Recommendations of restaurants in Taipei in a mobile device are
performed.
Woerndl et al. [161] tries to incorporate contexts in recommender systems to make
it applicable in mobile domain. The approach helps users to get aware of what have
been installed in mobile of their neighbor; accordingly they may get recommendations
for their mobile.
Though there are various techniques which have been classified as a separate
category of RS, however, we classify the following different recommendation
procedure as a part of context aware recommendation. Somehow, these are the
contexts which may affect the recommendations from both seller and buyer point of
view.
Location aware RS
Trust aware RS
Temporal RS
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
47 | P a g e
Brunato and Battiti[162] realized the need of pilgrims and suggested mobility-
aware recommendation system by fetching the location of the users. The authors have
calculated a preference metric which answers the queries of the users for their needs
of resources while making any pilgrimage. Mobility scenarios are introduced to better
appropriate and more reliable predictions of user requirements.
Levandoski et al. [163] utilized location based ratings for recommendation and
presented „LARS‟, a location-aware recommender system. User partitioning is used to
explore the location based ratings. The technique produces quality recommendation;
in addition it maximizes the scalability of the system [164].
Yang et al. [165] has identified the need of the customers and sellers both for
promotional selling and has presented a location based recommender system for
online shopping which gives the best recommendation by fetching the sales and
promotions which are location dependent. Tang et al. [154] has presented a location
aware system which also incorporate collaborative techniques to produces QoS web
based services. The web recommendations are made based on collaboration of user‟s
locations.
A Bayesian Networks (BN) influenced map-based customized RS is proposed
[166]. The system utilizes contextual knowledge including location and time. The
contexts like weather and user request automatically collected from mobile devices
are used to recommend appropriate item to users which match to their preferences.
Temporal recommender systems are meant to recommend items for users when
time is required to be kept an essential component in decision-making process. A
system is designed to recommend ranked cafes to customers [167] according to their
preferences, explored by their preference‟s knowledge, characteristics of the cafes‟,
specific situations, requirements, as well as the time of intended recommendation.
Queue Lee et al. [168] suggested a collaborative filtering-based recommender
system using implicit feedback. Since the system does not use explicit feedback, it
had relied upon pseudo rating observed from implicit feedback. The time of user‟s
purchase and launch of an item are used to construct pseudo rating matrixes which in
turn increase recommendation accuracy.
Lathia et al. [169] have shown how the temporal diversity can affect the
recommendation specially the behavior of CF techniques in recommendations. Since
the user‟s rating serve as a base in CF techniques. It is shown in their work that CF
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
48 | P a g e
data changes over time and a user may not always rate the item each time he/she
comes to shop online.
The authors in [170] presented a hybrid recommender system that not only
incorporates the demographic details of users but also the temporal information. The
results of the experiments has supported that temporal knowledge may enhance the
performance.
2.3.7 Social Network based Recommender Systems
The detail of the RS applied over social networking environment has been extensively
studied and presented by Zhou et al. [53]. The authors have tried to explore the pros
and cons and the opportunities of social network based RS.
An overview of the Foafing the Music system is presented [14], [171]. The system
used the text from RDF Site Summary (RSS) and Friend of a Friend (FOAF). The
Foafing based system predicts music to a user that matches to his essence of music
listening. Music information is collected from RSS feeds, music related blogs,
upcoming albums and „mp3‟ audio files at different music containing sites. The
system discovered music with the help of user profiling, information and descriptions
based on context supported ontological details of music domain.
Hu has presented a new paradigm of recommender systems. The RS can make use
of social networks (SN) based information. This information can be the preferences
observed for users, usual inclination of users towards a product or service, influenced
and influencing entities, like friends and acquaintances. A probabilistic model is
designed for personalization of the suggestion from these inferences. The real data
from online SNs are extracted. With their experiment, the author has concluded that
there is a strong similarity in the preferences of friends. Experimental results on this
dataset show that proposed system improves the performance [172].
To make use of social network where private or personalized data of an individual
is easily accessible for recommendation of items to new users are presented [136].
Human personality characteristics are integrated with rating given by them in the
recommendation process.
A social network based recommender system which exploit [173] the trading
relationships has been proposed. The system proposes the ways to compute the degree
of recommendation for trusted online auction sellers. The authors have utilized
network structure which is formed by history the transaction performed by user.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
49 | P a g e
2.3.8 Soft Computing Techniques based Recommender Systems
The soft computing techniques have now been increasingly used in recommender
systems for incorporating collaborative recommendations, reclusive recommendations
and hybrid recommendations. To deal with the uncertainty in various business
marketing affairs, Cornelis et al. [174] make use of fuzzy relations to model the
degree of similitude between items and users. They also proposed a novel hybrid CF–
CB approach whose rationale is concisely summed up as “recommending future items
if they are similar to past items that similar users have liked”. A hybrid fuzzy logic-
based recommendation framework [175] was then developed to improve the trade
exhibition recommender system for e-government. Zhang et al.[176] has developed a
telecom recommender system using fuzzy techniques. The authors have used fuzzy on
item based similarity approaches. The have applied fuzzy set techniques on mobile
product and service recommendation. They have designed system referred as Fuzzy-
based Telecom Product Recommender System (FTCP-RS).
A soft computing technique is applied for the recommendation of the books for
university graduates by the authors [35] where they have incorporated the vagueness
in the preferences of the books and aggregated the score of the books using OWA
technique. Similar work has been suggested using ordered ranked approach in [5].
Hybrid approach using fuzzy-genetic to exploit the use of its varieties to address
sparsity and scalability is addressed. But CF techniques face the issue of accuracy and
sacability both. To overcome the problems of accuracy and scalability with memory
based and model based CF techniques, respectively. The proposed system reduces
sparsity and complexity; while retaining the neighbor recommendations perspective
[177].
Fuzzy logic based RS is presented as a solution to issues encountered by CF
techniques for some specific situations regarding those items which are brought into
market rarely and not necessarily be repetitively put on sale [175]. The employed
fuzzy technique recognizes the uncertainty in the information. The method can be
helpful in various scenarios like trade exhibition recommendation.
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
50 | P a g e
Table 2.3: Recommender Systems, Categories and Techniques
S.N
o
Types of
Recommende
r System (RS)
Sub-
category Techniques Research Papers
1
Collaborative
Filtering (CF)
based RS
Item
based
Association rule mining between
preferences of neighbor of users,
Rating, Choice of individuals for
varied items, Similarity in the
preferences of different users for
common items, Tagging.
[178], [42], [57], [148],
[179], [180], [181], [94],
[85], [96], [179], [182]–
[185], [8], [9], [30], [43],
[165]
User
based
Model
based
Bayesian networks, clustering,
Machine learning, Graph
modeling
[88], [68], [88], [41], [79],
[80],
[81][84][86][83][85][31][82
], [87].
2
Reclusive
Methods (RM)
based RS
Heuristic
method
Rule induction, nearest
neighborhood, Rocchio‟s
algorithm, tagging, rating, etc.
[98], [134], [178], [55],
[101], [187]–[191], [81],
[94],[95][96], [97].
Model
based
technique
s
Bayesian networks, clustering,
Machine learning, Graph
modeling
[189],[89], [192]–[194],
[195], [91], [196], [98],
[71], [99], [100]
,[101][102][103].
Web
mining
Opinion mining, web usage
mining, etc.
[104], [61], [105], [106],
[39], [38],
[40][107][1][78][108][96][1
5]
3
Hybrid
recommender
systems
CF
dominate
d RM
Techniques of CF, RM applied
with each other in different
combinations
[47], [197], [135], [48],
[198], [199], [200],
[201],[202], [203], [204],
[96], [145], [183]
RM
dominate
d CF
CF and
RM
coalesced
into one
Subseque
nt
Integratio
n of
separately
applied
CF and
RM
Integratio
n of CF
and RM
with
Techniques of CF and RM are
applied with KBS, and other
[205], [123], [206]–
[209], [202], [19], [8],
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
51 | P a g e
(KBS) fuzzy, social network, etc. [183], [197], [199],
Integratio
n of CF
with other
than RM
Integratio
n of RM
with other
than CF
4
Demographic
filtering based
RS
- Correlation, similarity measures,
etc.
[109], [209] , [210], [211],
[109], [212], [71],
[110],[213], [214]
5
Knowledge
based
Recommender
System (KBS)
Constraint
based Machine learning, Bayesian
network, AI, etc.
[121], [215], [125],
[216], [126], [217], [122],
[218], [127], [128] , [111],
[114], [219]–[221], [120],
[7], [222], [220], [119]
Case
based
6
Context Aware
Recommender
System
Location
aware,
Temporal,
Trust
aware
User feedback, AI
techniques, machine learning,
etc.
[223], [224], [225],
[226], [161], [8], [163],
[154], [164], [227], [228],
[167], [170], [168], [155],
[229]–[231]
7 Social network
based RS
Foafing,
trade
relationsh
ip, etc.
Similarities measures, user
profiling, etc.
[232], [233], [234], [18],
[172]
8
Soft
Computing
techniques
based RS
Fuzzy
genetics,
fuzzy
linguistics
,
OWA, ORWA, fuzzy model, etc. [5], [35], [235], [174],
[236], [237], [25], [177]
The problem with CF technique and RM is that both of them fail in representation
of explanation of relationship between users feedback and features of items as they
are subjective and uncertain. The authors have presented Fuzzy set theoretic method
(FTM) [237] which identifies the application of fuzzy method presented by
Yager[89]. The FTM makes use of aggregation which finds confidence score for
recommendation. The techniques also utilize the various statistical measures to
evaluate the RS.
The authors have suggested how to automatically recommend newly launched
items to user which have no prior rating. Only with the users past history of purchase,
the new items are recommended to users. The combination of Bayesian networks and
Fuzzy Set Theory are used to enhance the system performance [238].
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
52 | P a g e
2.4 Summary
The comprehensive survey of the recommender system is presented in this chapter.
With the help of the study conducted in the chapter, we have concluded that there is
an exponential growth of the research in the field of RS. Researchers have shown a
great interest towards this area. The application area of RS has covered diverse field
of daily life. It includes academia, health and care, business using e-commerce and e-
shopping sites, etc.
Various techniques have been used to meet the demand for these applications. We
have categorized 8 different types of RS which is further broken into 19 sub
categories based on techniques and filtering algorithms used. The Collaborative
Filtering (CF), most influential recommender technique, has largely used by the
researchers but still fails to produce satisfying solution due to major drawbacks like
cold start problem for new users and sparsity, as stated in section 2.4. The leading
technique next to CF widely used in the literature is „Reclusive Method‟ (RM) or
„Content based filtering‟; the technique also suffers from the same complications. No
technique alone can sensibly be considered as a solution to these problems, instead
hybrid approach may fulfill the requirement. Thus, a more robust hybrid method
which incorporates the best of these techniques without being affected by their worst
may produce satisfying results.
It is evident that prior to designing a recommender system one must understand the
characteristics of the recommendation which can please the users. User‟s feedback
directly reflects their priorities, likes and dislikes. Therefore, explicit or implicit
feedback from the users to know the characteristics of their past preferences as well as
to predict the future behavior is pragmatically important. A recommendation
procedure for the books will be extensively discussed in Chapter 3 which exploits the
feedback from users (experts from the universities are considered as users). The
procedure provides a recommendation on consensus basis which overcomes the
prevailing issues with RM and CF techniques.
In section 2.3.8, it is discussed how the soft computing have emerged and its
employment in recommendation technology is seen increasing rapidly. In Chapter 4 a
Chapter 2: An Overview of Recommender Systems
___________________________________________________________________________________
53 | P a g e
method is discussed which utilizes soft computing techniques and makes possible use
of it to outstretch the satisfaction level of users.
Also, it is revealed that rating scale proves handicap for several occasion, like
when no ratings are available or rating scale lacks standard. Opinion mining could
help better recommendation where rating scale might not have done well. Opinion
mining could provide the recommendation by finding user‟s requirement according to
their reviews, and matching it with characteristic of the product, hence,
recommending the exact items to users. An opinion mining based recommendation is
also proposed and described in the Chapter 5. The proffered approach is believed to
be a realistic one adequate for the users‟ satisfactions.
54 | P a g e
Chapter 3
Link Mining based Book Recommendation
Approach
3.1 Introduction
The development of modern tools and technologies, and inclination of the new
generation towards education has made the demand of Information exchange very
high. The huge information available on a specific topic creates confusion for the
people who are seeking for desired source to grab the correct information. Books and
research articles, whether online or offline, are the sources for obtaining information.
Hence, it is an important task to filter the sources for finding the desired books. There
are millions of e-books available on the Internet and the numbers are increasing
rapidly, this rapid increase has created a high demand of developing a
recommendation technique to get exact and desired book.
There are a good number of works in the area of product recommendation [2].
There are various methods being used frequently in recommendation techniques.
Collaborative filtering and Content-based Recommendation are most frequently used
recommendation techniques found in the literature. Due to some serious problems that
collaborative filtering faces, researchers switched to Web mining techniques for
product recommendation problems. Web usage mining; a process of extraction of
useful patterns from web usage data, supposed to be the most applied branch of the
web mining techniques that has attracted the researchers in recent decade [239].
Generally recommender systems use customer‟s preferences and assume that
several customers must have same taste and may like the similar products, however it
is not the case always. If we are concerned about academic books, the selection is not
the fun and must not be dependent upon student‟s choice, but it should be handled
with utmost care and decided by the experts. Therefore specialized way of
recommendation where authorities‟ recommendations are considered would be
advisable and fruitful. Therefore it seems adequate while recommending books that
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
55 | P a g e
one should ask the percept of experts at Universities instead of common people so that
the above issues can be avoided.
Thus, instead of applying personalized recommendation approach it seems
adequate to make use of a group recommendation technology and same techniques
and single ranked recommendation can be one answer to several simultaneous queries
[31]. With the above discussions in the considerations, we have suggested a ranked
recommendation approach for books which aggregates the several ranking of the top
universities (which is considered as authorities) and employ link mining approach in
the recommendation process. On the one hand it handles the cold start issues and on
the other hand it eases the complexities of personalized recommendation to huge
number of users and replaces it with a single ranked recommendation.
In this chapter, we have recommended top books for University‟s students in
Indian scenario. That is why we have chosen top ranked universities from India. The
selection of top ranked universities is based upon the QS world university ranking. QS
Ranking is one of the leading ranker of the academic institution. Once the top
institutions are explored, their syllabus for the particular subject are searched which
served as a base for the recommendation of books.
Figure 3.1: An overview of link mining approach
Web page A
Web page B
Web page E
Web page I
Web page H
Web page G
Web page F
Web page C
Backward Link
Forward Link
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
56 | P a g e
Positional Aggregation Scoring (PAS) technique [34] is highly advisable for
aggregating a final result from these types of data. Instead of searching desired books
for a user from thousands of books available, it seems more appealing to find the best
amongst the books prescribed by ranked university, as it shortens the data overload,
reduces the complexities and increases the authenticity of the products.
We have chosen top universities amongst the Indian universities and checked their
recommendation for different courses of computer science; it‟s evident that a
recommendation of a book by a high class university will eventually increase the
importance of the recommended books. In this way the philosophy of link mining is
incorporated in this chapter. Apart from PAS technique, there are some other useful
aggregation operators which can be proved effective. The primary advantage of the
adopted technique is that it includes the recommendation of high status top ranked
universities as well as rank of the rankers i.e. universities, which are authorities for
the academic program to recommend books for university students. Figure 3.1depicts
an example of link mining.
The main contributions in the Chapters are as follows:
We have conferred a book recommendation method based on aggregation of
expert‟s decision. Thus the problem of book recommendation is converted into
decision making problem.
Positional Aggregation based Scoring Technique is implemented for
recommendation process. To the best of our knowledge, we are the first to use
this concept for book recommendation.
Section 3.3.1 deals with the details of data collections, types of dataset and
experimental results, in which there is a comprehensive discussion to highlight the
various aspects of experimental results. Finally we have summarized in section 3.4.
3.2 Book Recommendation using Positional Aggregation based Scoring
Technique
We are concerned with different books prescribed in the syllabus of top ranked
universities. The prescribed books may be considered as the rankings of the books by
that particular university. As discussed in section 3.5.1, the syllabus of respective
universities differs significantly. Thus, it gives us a partial list for the books
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
57 | P a g e
recommended in the syllabus by different universities. We have aggregated the ranked
books to obtain a full list of books with aggregated ranking. For full list, we have
several well-known methods like Borda‟s method [240] Markov chain based methods
[241]and soft computing based methods. But these techniques work for full list only.
[240], [242]. Therefore we have applied Positional Aggregation Score (PAS) based
technique [34], [243] that can work better to recommend the top books for partial list.
3.2.1 Positional Aggregation based Scoring Technique
The PAS based technique is used to aggregate the ranked data which have been
ranked by several users, hence, involves different position in the data set for different
ranking. R. Ali [243]has used it to evaluate a RS designed for product
recommendation. The PAS technique works as follows:
Suppose „m‟ different books are recommended by „n‟ different universities. First
we find out the rank of a book „Bi‟ for every university, we assign maximum value
(Vmax= -1) to book which is best ranked i.e. first ranked book is assigned a value '-1'.
The idea behind assigning „-1‟ to best ranked book is to give highest value to it and all
the values associated to ranked book should be in order of their ranking i.e. better
ranked books must have a higher numerical value associated with it. For next value,
we assign {(Vmax) – (i)} to (i+1)th
best ranked book. The above steps are repeated so
that all the books are assigned a value. If a book is not ranked, it is assigned a value „–
(m+1)‟, where m is number of total books being ranked by different universities.
Now, we compare each book „Bi‟ with all the „m-1‟ books. If value of a book is
greater than the other, we assign value of Bi = 1 otherwise Bi is assigned 0, i.e Bi = 0;
If it is found that Bi == - (m+1), again zero is assigned to Bi. Finally sum of all values
of Bi for each university are obtained. Thus, we will be getting (m+1) different scores
of every book; we call it S. the final score „FS‟ is given by (S / (m+1)).
3.2.2 Book Recommendation Approach using Positional Aggregation Scoring
The working of PAS is described in section 3.2.1. The above procedure is used for
recommending books. The following example is illustrates the approach in detail.
Example 3.1: we are taking an example with four books and five universities to
illustrate the above procedure, i.e. m = 4. U_1, U_2, U_3, U_4 and U_5 are five
different universities and B1, B2, B3, and B4 are four different books. Sequence 1, 2, 3
and 4 are ranking position of the books, i.e. the row consists of „1‟ in first column will
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
58 | P a g e
give the first ranked book of the particular university. All the universities have their
own ranking for different books; these rankings are given in Table 3.1.
Let a cell „z‟ is represented by z (r, c), where r and c represent the rth
row and cth
column respectively. The value of z (1, 1) is B1, i.e. B1 is ranked first by University
U_1. z (3, 5) is „-„ that implies no book is recommended by University U_4 except B2
in first four position. In Table 3.2, rank to score conversion is illustrated where the
best ranked university is assigned „-1‟. As B1 is first ranked book by University U_1,
the cell corresponding to B1 and S (U_1) has -1, where S (U_1) denotes score
assigned to respective books by U_1. Those books which are not ranked by any
university is assigned a value „-5‟.
Table 3.3gives the pairwise comparison of respective books. A pair (Bi, Bj) = 1
implies that book Bi is preferred over book Bj by the university. If (Bi, Bj) = 0, it
means book Bj is preferred over book Bi by the university concerned. The column
PwC (U_1) implies pairwise comparison of the books for U_1. The sum of values of
all the comparison of each book for all the universities is shown in Table 3.6. This
value is termed as preference score, and hence the notion for this in first column is PS
(U_1), i.e. preference of a book by U_1, and so on.
Table 3.1: Top 4 ranked books by 5 universities
Rank positions U_1 U_2 U_3 U_4 U_5
1 B1 B2 B3 B2 B4
2 B2 B4 B1 - B1
3 B3 - B4 - -
4 - - - - -
Table 3.2: :Conversion of Rank into Scores
S (U_1) S (U_2) S (U_3) S (U_4) S (U_5)
B1 -1 -5 -2 -5 -2
B2 -2 -1 -5 -1 -5
B3 -3 -5 -1 -5 -5
B4 -5 -2 -3 -5 -1
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
59 | P a g e
Table 3.3: :Pairwise comparison of books
Pair (Bi, Bj) PwC (U_1) PwC (U_2) PwC (U_3) PwC (U_4) PwC (U_5)
B1,B2 1 0 1 0 1
B1,B3 1 0 0 0 1
B1,B4 1 0 1 0 0
B2,B1 0 1 0 1 0
B2,B3 1 1 0 1 0
B2,B4 1 1 0 1 0
B3,B1 0 0 1 0 0
B3,B2 0 0 1 0 0
B3,B4 1 0 1 0 0
B4,B1 0 1 0 0 1
B4,B2 0 0 1 0 1
B4,B3 0 1 0 0 1
The value of B1 for U_1 comes out to be 3, which means university „U_1‟ prefers
book „B1‟ over rest of the three books. There are (m-1) comparisons for each book;
hence the values obtained in Table 3.6 can be normalized by dividing 3.
Table 3.4: Normalized preference score of books
NPS (U_1) NPS (U_2) NPS (U_3) NPS (U_4) NPS (U_5)
B1 1 0 0.66 0 0.66
B2 0.66 1 0 1 0
B3 0.33 0 1 0 0
B4 0 0.66 0.33 0 1
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
60 | P a g e
Algorithm 3.1: Positional Aggregation based Scoring of books
𝑷𝑨𝑺 = 𝟏
𝒏 𝑵𝑴 𝒊,𝒌
𝒎,𝒏
𝒊,𝒌=𝟏
Preliminaries:
Total no .of books is „m‟
Total no. of different universities is‟ n‟,
hence total „n‟ ranking is available
For each book Bi belongs to m, we have
different ranked position of Bi in every
ranking Rk ; 1≤k≤n i.e. we have a matrix
with m rows and n column may be
represented as: R[i,k] where 1≤i≤m
&1≤k≤n;
Steps:
I: Repeat the following procedure till steps7 for every
ranking Rk
1: find out the rank of a book „Bi‟ where, 1≤i
≤m
2: Assign maximum value (Vmax= -1) to book
which is best ranked
3: For next value, assign (Vmax – i) to (i+1)th
best ranked book
4: If a book is not ranked, assign it a value = –
(m+1)
Repeat the steps 2 to 4 until all the books are
assigned a value, store these values in a matrix SM [i,
k];
5: compare each book „Bi‟ with each of the
remaining „m-1‟ books, for 1≤i≤m
If
SM [i,k] > SM [j,k], 1≤ j ≤ m; i ≠ j
PC [i,k] = 1;
else
PC [i,k] =0;
6: we find preference score matrix PSM [i,k]
such that
PSM [i,k]= 𝐁𝐢,𝐁𝐣 𝒎𝒋=𝟏 ; j≠i
7: create normalized matrix NM [i,k] = PSM
[i,k] /(m-1);
8: we find positional aggregation based scores
as:
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
61 | P a g e
Figure 3.2: Positional Aggregation Scoring based Book Recommendation System
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
62 | P a g e
Table 3.5: Positional Aggregated scores of books
Book Positional Aggregation Score
B1 0.464
B2 0.532
B3 0.266
B4 0.398
Table 3.6: Preference score of books
PS (U_1) PS (U_2) PS (U_3) PS (U_4) PS (U_5)
B1 3 0 2 0 2
B2 2 3 0 3 0
B3 1 0 3 0 0
B4 0 2 1 0 3
Table 3.7: Ranked books based on Positional Aggregation Scoring technique
Ranked position Book
1 B2
2 B1
3 B4
4 B3
The normalized score is given in Table 3.4. We call it normalized preference score.
Finally we get Positional Aggregated Score by dividing the values obtained in Table
3.4 by number of university „n‟, here n=5. The values are given in Table 3.5.
Finally the PAS is sorted to find top books, as shown in the Table 3.7. The above
calculation is summarized in algorithm 3.1. Pictorial representation of the working of
book recommendation using PAS based technique is presented in Figure 3.2.
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
63 | P a g e
3.3 Results and Discussions
In the previous sections of this chapter, we have described the procedure of
recommendation and generalized steps for the implementation of the recommendation
scheme is presented in respective sections. In this section, the results of the
recommendation by different techniques applied in the previous discussions, are
presented and discussed in details. The methods and steps of data collections, data
filtering, datasets, and pros and cons of the methods are also discussed.
3.3.1 Dataset
Basically, we are concerned in recommending books for university graduates of
Indian universities. Initially, those different books were taken that could be a part of
the curriculum of the universities. Though, there were neither any criteria nor any
limit of the inclusion of the books. Then we had a second thought to filter the data in a
way that could fulfil our objective, i.e. top ranked books of the course to the students.
For this, it seems adequate to include the top ranked institutions and their
recommended books. This step will filter the data as well as it helps in explaining why
the methodology (section 3.4) is chosen for the recommendation process. Also, only
„computer science‟ as a subject is selected from these universities/institutions.
Because, once we can find the method of presenting top books for any specific
subject, it can easily be extended for all other subjects. Hence, different courses of
computer science like, Discrete Mathematics, Data Structure, etc. which are almost
considered in top institutions, have been included. The methods of selection of
universities and courses, and the final recommendation by the different techniques are
discussed in subsequent sub sections.
3.3.1.1 Selection of top Universities
The selection of top universities can fulfill the purpose of recommendation of
appropriate books for graduate students as far as our methodology is concerned.
Although, the basic idea seems to be simple, however, is very useful. The books
which are recommended by top institution should be more reliable for students rather
than the books recommended by other corporate recommender sites. Thus, we have
chosen top universities from India and their recommended books are explored. The
recommended books of these universities are quantified and then aggregated using
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
64 | P a g e
proposed technique with the help of some strong aggregating operators which gives
top N books to be suggested.
Table 3.8: Top 7 Indian Universities in QS ranking [244]
Rank Position University Name
1 Indian Institute of Technology, Bombay
2 Indian Institute of Technology, Delhi
3 Indian Institute of Technology, Kanpur
4 Indian Institute of Technology, Madras
5 Indian Institute of Science, Bangalore
6 Indian Institute of Technology, Kharagpur
7 Indian Institute of Technology, Roorkee
There are several ranking sites and authorities which suggest top universities,
courses and places for higher studies. Amongst all these rankers, QS World
University rankings in collaboration with Elsevier is one of the leading and highly
reliable source of university ranking that ranges about 40 subjects all around the
globe. They ranked the institutions subject-wise, region-wise, course-wise, etc. QS
World University Rankings for computer science & information systemshas only 7
Institutions from India[244]. We have selected these institutions for inclusion of
books in our procedure. The list of top Indian Institution for 2015 is listed below in
Table 3.8.
3.3.1.2 Courses included from top Universities
The different institutions have offered different courses for their enrolled students.
Merely, there are the very few courses having exactly the same title at these top
institutions. However, we have tried to our best for exploiting the courses that can be
categorized as a common to these institutions. Although, there are the subjects that
differ in these universities and academic institution, we have just taken those courses
which are either common to all or most of the universities offer these courses and
have published its syllabus at respective websites, or have made these syllabus
available to their students. The list of the courses included in the proposed work, and
corresponding offering universities is given in the Table 3.9. The column representing
U_1, U_2 to U_7 is the sorted universities, i.e. U_1 is the best ranked
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
65 | P a g e
Table 3.9: Syllabus of various Courses, offered at top Universities.
Sequence Course Title U_1 U_2 U_3 U_4 U_5 U_6 U_7
1. Discrete Mathematics
2. Operating Systems
3. Theory of Computation
4. Computer Networks
5. Software Engineering
6. Compiler Design
7. Principles of Database Systems
8. Artificial Intelligence
9. Data Structure
10. Graphics
university. Discrete Mathematics is offered at every university, whereas, only ranked
3 university does not offer it. More specifically, except IIT Kanpur all other top
universities offer the course with similar title, “Discrete Mathematics”. In the same
way, the table depicts the institutes and whether the courses are offered their or not.
The „‟ and „‟ marks represent syllabus are available or not for the courses with the
same title at corresponding universities, respectively. However, few universities do
not show the entire syllabus at their web site. We had to make an extra effort by
visiting the faculty and students of the institution. Some of them are contacted in
person, while others are being asked by email, etc.
3.3.1.3 Prescribed books by top Universities:
10 courses are listed in the Table 3.9. The universities recommended the syllabus
where they explicitly describe the prescribed books for students. Different courses
have different number of books. The list of courses and their respective books,
obtained by the combining all the universities are listed in Table 3.10.
In the above table, total number of books is given. Discrete Mathematics,
represented by code „DM‟, has a total of 17 books from all the selected top 7
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
66 | P a g e
Table 3.10: Total number of books in the syllabus of corresponding courses in top Universities
Sequence Course Title Course Code Number of
books
1. Discrete Mathematics DM 17
2. Operating Systems OS 15
3. Theory of Computation ToC 11
4. Computer Networks CN 18
5. Software Engineering SE 19
6. Compiler Design CD 10
7. Principles of Database
Systems
DB 13
8. Artificial Intelligence AI 20
9. Data Structure DS 15
10. Computer Graphics CG 20
universities. It is also evident from Table 3.13 that only six universities have shown
the details of the syllabus for the book „DM‟ at their websites. Thus, these 17 books
are from 6 universities. In the same way, the total books for all courses are the
collection from universities which have made their syllabus available, whose
detailsare presented in Table 3.13. Maximum number of books considered is of
Artificial Intelligence „AI‟ and Computer Graphics „CG‟. Both courses have total
collection of 20 books from these universities. However, not all 7 universities have
syllabus for both the books published at their websites. There are six and five
universities involved for both the books respectively. Minimum number of books
available is of Compiler Design „CD‟.
Only 10 books are available, although 5 universities have made their syllabus
available.But the number is low because most of the recommended books are
common to these universities, whereas in the case of CG, there is significant
difference in recommended books. Thus 158 different books for 10 different courses
are included in our procedure.
3.3.2 Experimental Results
In this section, we have discussed the final recommendations by all 5 different
techniques implemented which include PAS, OWA with quantifiers „at least half‟, „as
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
67 | P a g e
many as possible‟, „most‟ and ORWA. We have applied all the techniques which are
discussed above.
Different ranked books are obtained by using all these techniques. The books are
represented by unique course codes. E.g.: code „CD‟ has been used to refer books on
Compiler Design see table (3.14). In the similar way, different courses of books have
different notation for representations. For each book we have different sequence of
the books according to their ranking. For „Compiler Design‟ the different codes are
CD1, CD2, etc. The details of the books on compiler design including code, author of
the books, title and publisher, for which syllabus are available, are listed in Table
3.11.Course code of the books and corresponding rank given by respective
universities are given in Table 3.12.
Table 3.11: Code and details for books on Compiler Design
Course Code Author Title
Publisher
CD.1.
Alfred V. Aho, Monica
S. Lam, Ravi Sethi and
Jeffrey D. Ullman:
Compilers: Principles,
Techniques, and Tools 2/E, AddisonWesley
2007.
CD.2. Andrew Appel
Modern Compiler
Implementation in
C/ML/Java
Cambridge University
Press, 2004
CD.3.
Dick Grune, Henri E.
Bal, Cerial J.H.
Jacobs and Koen G.
Langendoen:
Modern Compiler
Design
John Wiley& Sons,
Inc. 2012.
CD.4. S. Muchnick,
Advanced Compiler
Design &
Implementation,
Indian Reprint 2002.
CD.5. K Cooper,
L Torczon
Engineering a Compile
r
2nd Ed., Morgan Kauf
mann, 2011
CD.6. KC Louden, Compiler Construction:
Principles and Practice
Cengage Learning, 199
7
CD.7. D Grune et al.
Modern
Compiler Design Wiley, 2000
CD.8. Michael L Scott, Programming Languag
e Pragmatics
3rd Ed., Morgan Kauf
mann, 2009
CD.9.
Tremblay, J.P. and
Sorenson, P.G.
Theory and Practice of
Compiler Writing
SR Publications.
2005
CD.10.
Tremblay, J.P. and
Sorenson, P.G.
Parsing Techniques: A
Practical Guide
Ellis Horwood.
1998.
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
68 | P a g e
Table 3.12: Ranked list of book „compiler design‟ by top universities
Rank
Position U_1 U_2 U_3 U_4 U_5 U_6 U_7
1. CD1 CD2 CD1 CD1 CD1 - CD1
2. CD2 CD4 CD2 - CD4 - CD9
3. CD3 CD1 CD5 - - - CD5
4. - - CD6 - - - CD6
5. - - CD7 - - - CD10
6. - - CD8 - - - -
7. - - CD4 - - - -
8. - - - - - - -
9. - - - - - - -
10. - - - - - - -
From the table it is evident that U_1 has ranked book CD1 1st. CD2 and CD3 are
ranked 2nd
and 3rd
respectively. As U_1 is the 1st ranked university, this implies that
top university has recommended only three books on compiler design for their
students. The book CD1 is almost ranked by all the university except U_6, which has
not issued list of any book for the particular course. U_2 has ranked CD1 3rd
and CD2
is ranked 1st. However, U_3, U_4, U_5 and U_7 all have ranked CD1 1
st. It is also
observed that U_4 has recommended only one book. The PAS based ranking of books
are obtained by applying the procedure illustrated in example 3.1. To get the ranking
of books, say compiler design, we have considered all 10 books recommended by top
universities. These ranks are numerically represented in Table 3.13. In the table, R
(U_1) indicates rank given by U_1. These ranks are converted into scores and shown
in Table 3.14. As described in Table 3.2 of example 3.1, the best rank i.e. „1‟ is
assigned „-1‟, rank 2 is assigned „-2‟, and so on. The book which is not ranked by any
of the university is assigned lowest value which is „-8‟ here. Thus cells values
„0‟ofTable 3.13have changed to „-8‟ in Table 3.14.
Table 3.13: Compiler design ranked books by top 7 Universities
Course
Code R (U_1) R (U_2) R (U_3) R (U_4) R (U_5) R (U_6) R (U_7)
CD.1. 1 3 1 1 1 0 1
CD.2. 2 1 2 0 0 0 0
CD.3. 3 0 0 0 0 0 0
CD.4. 0 2 7 0 2 0 0
CD.5. 0 0 3 0 0 0 3
CD.6. 0 0 4 0 0 0 4
CD.7. 0 0 5 0 0 0 0
CD.8. 0 0 6 0 0 0 0
CD.9. 0 0 0 0 0 0 2
CD.10. 0 0 0 0 0 0 5
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
69 | P a g e
Table 3.14: Rank to Score conversion of book Compiler Design
Course
Code S (U_1) S (U_2) S (U_3) S (U_4) S (U_5) S (U_6) S (U_7)
CD.1. -1 -3 -1 -1 -1 -8 -1
CD.2. -2 -1 -2 -8 -8 -8 -8
CD.3. -3 -8 -8 -8 -8 -8 -8
CD.4. -8 -2 -7 -8 -2 -8 -8
CD.5. -8 -8 -3 -8 -8 -8 -3
CD.6. -8 -8 -4 -8 -8 -8 -4
CD.7. -8 -8 -5 -8 -8 -8 -8
CD.8. -8 -8 -6 -8 -8 -8 -8
CD.9. -8 -8 -8 -8 -8 -8 -2
CD.10. -8 -8 -8 -8 -8 -8 -5
Table 3.15: Positional Score for book Compiler Design
Course
Code
PS
(U_1)
PS
(U_2)
PS
(U_3) PS (U_4)
PS
(U_5)
PS
(U_6)
PS
(U_7) PAS
CD.1. 1 0.777 1 1 1 0 1 0.8252
CD.2. 0.888 1 0.888 0 0 0 0 0.3965
CD.3. 0.777 0 0 0 0 0 0 0.111
CD.4. 0 0.888 0.333 0 0.888 0 0 0.3012
CD.5. 0 0 0.777 0 0 0 0.777 0.222
CD.6. 0 0 0.666 0 0 0 0.666 0.1902
CD.7. 0 0 0.555 0 0 0 0 0.0792
CD.8. 0 0 0.444 0 0 0 0 0.0634
CD.9. 0 0 0 0 0 0 0.888 0.1268
CD.10. 0 0 0 0 0 0 0.555 0.0792
Table 3.16: Ranking of book „compiler design‟ using Positional Aggregation Scoring
Rank position PAS based ranking
1 CD.1.
2 CD.2.
3 CD.4.
4 CD.5.
5 CD.6.
6 CD.9.
7 CD.3.
8 CD.7.
9 CD.10.
10 CD.8.
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
70 | P a g e
Table 3.17: PAS based Ranking of different books
Rank
Position
PAS based Ranking of different books
1 DM.9. AI.2. DS.1. DB.2. CG.2. SE.2. OS.2. CN.2. TOC.1.
2 DM.1. AI.5. DS.2. DB.1. CG.7. SE.4. OS.4. CN.17. TOC.2.
3 DM.10. AI.21. DS.4. DB.4. CG.1. SE.9. OS.3. CN.1. TOC.6.
4 DM.8. AI.7. DS.6. DB.7. CG.3. SE.3. OS.5. CN.12. TOC.4.
5 DM.2. AI.11. DS.12
DB.3. CG.14. SE.18 OS.13. CN.7. TOC.3.
6 DM.4. AI.4. DS.3. DB.5. CG.16. SE.1. OS.1. CN.11. TOC.5.
7 DM.12. AI.6. DS.7. DB.6. CG.4. SE.11 OS.10. CN.13. TOC.7.
8 DM.16. AI.1. DS.5. DB.8. CG.17. SE.5. OS.7. CN.8. TOC.9.
9 DM.3. AI.14. DS.8. DB.9. CG.5. SE.8. OS.11. CN.14. TOC.8.
10 DM.5. AI.15. DS.13
DB.13. CG.15. SE.12 OS.8. CN.18. TOC.10.
The final PAS is given in Table 3.15. Thesescores obtained by PAS lead to the
ranking of books on different disciplines. The ranking of books on compiler design is
given in Table 3.16,the CD1 has almost attained maximum score from all the adopted
methods. Its score by PAS is 0.8252.
In the ranking of books on „Discrete Mathematics‟, DM.9 is placed at 1st position
and DM.1 is placed at second position. However, DM.10 is 3rd
ranked and DM.3 is
ranked at 9th
positions. AI.1 is not even in top 3 ranking and is ranked 8th
whereas
AI.2 is at top position. If we observe the rankings of data structure books, the top 2
position is acquired by first two books ranked by top universities, i.e. DS 1 and DS 2.
This variation indicates the importance of the authoritative recommendations. In most
of the cases it is observed that universities‟ recommendation has influenced the
recommendation significantly.
The books on data base has the same trend of being ranked by the PAS technique
and on observing the rankings of data base books, the top 2 position is acquired by
first two books ranked by top universities, i.e. DB. 1 and DB. 2. Similarly, computer
graphics (C.G) books are ranked and has similar variations as of data structure books,
however, top 3 books which is recommended by the authoritative recommendation are
also in top 4 ranking position of PAS. But beyond top 10 positions in the ranking of
universities, three books have also secured ranking in PAS under 10 ranks. The books
on Software Engineering interestingly has not ranked SE.1 in any of the top 5
positions, however SE. 2 is ranked top. Again, the books on Operating Systems
Chapter 3: Link Mining based Book Recommendation Approach
___________________________________________________________________________________
71 | P a g e
interestingly has not ranked OS.1 in any of the top 5 positions, however OS. 2 is
ranked top.
Unlike the above two books‟ ranking, the ranking of books on Computer Networks
has CN. 1 and CN.2 on top three positions which again suggest he importance of
authorities in ranking.
3.4 Summary
We have incorporated link mining approach for the recommendation of books. The
syllabus of the top ranked universities are taken and aggregated ranking which is
obtained by considering the most valuable universities‟ recommendation as more
preferred, with the help of positional aggregation scheme.
Rank aggregation algorithm which is termed as Positional Aggregation based
Scoring (PAS) technique is used for recommendation of books. We believe the
proposed technique may meet the user‟s need and provide them the perfect books they
need. For the sake of illustration and ease of experiments, we have shown the
procedure considering books from top institutions. However, we can generalize the
procedure of the recommendation for any kind of items. The robustness of the
procedure may lead to a novel way in the field of recommendation and would fulfill
the demand of millions.
72 | P a g e
Chapter 4
Book Recommendation based on Soft
Computing Approaches
4.1 Introduction
The Internet is the gift of the modern era, which is a consequence of the proliferation
in the modern technologies. The growth of the Internet has also boosted ecommerce.
Online shopping has become much more popular. Today it is a vogue for a common
man to shop online using online marketing portals such as www.amazon.com. The
boom in the Internet has caused data overload over it. The huge data over the World
Wide Web has increased the problems for the users to extract the exact information.
The buyer finds it extremely tough to go for an exact product which he or she is
looking for. While browsing the online shopping portals, multiple options are weeded;
however picking the right item is an arduous job. Researchers have proposed different
recommendation techniques to help the customers in purchasing the right item.
Various efforts have been made for an ease and effective online shopping. In last few
years, researchers have proposed a good number of recommendation techniques [2],
[3], [4]. An increase in soft computing methods in the field of recommendation
technologies especially, fuzzy based recommendation is recorded [177]. To solve the
various issues encountered in leading recommender systems, researchers have also
used web mining, an emerging recommendation technique that researchers are using
frequently. Link mining, supposed to be a sub set of the web mining technique, is an
emerging research area [245], [246].
In this chapter, we have proposed a recommendation methodology that
incorporates soft computing techniques and link mining both to rank the products and
recommend it before the users. Like the previous chapter, we have included the
importance of the recommending universities i.e. important links are given priority to
incorporate the link mining techniques. The soft computing methods are applied over
the link mining for a precise and near to preference recommendation for users.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
73 | P a g e
The two soft computing based averaging mechanism, Ordered Weighted
Aggregation (OWA) and Ordered Ranked Weighted Aggregation (ORWA), a
modified OWA, [5], [35] are applied. OWA is a fuzzy based averaging operator,
which in combination with linguistic quantifier gives a variation of option for decision
making problems. However, there is a lack of consideration for the voters or rankers.
By voters or rankers we mean the entity which recommends or suggests the items in
consideration. In our case, the item is book. The motive behind introducing ORWA is
to add the value of rankers in support of the philosophy that recommendation by
higher authority must be valuable than by a lower authority. In ORWA, a specific
weight is assigned to ranking agents (rankers), in our case; the universities. The
weight assignment method gives high weightage to the best ranked university and
hence their rankings are evaluated with a high degree of preference than to those
institutions that have lower ranking. The primary advantage of the adopted technique
is that it includes the recommendation of high status top ranked universities as well as
rank of the rankers i.e. universities, which are authorities for the academic program to
recommend books for university students.
On the one hand OWA utilizes different linguistic quantifiers to overcome the
problem of vagueness in the recommendation, and on the other hand ORWA
considers top ranked universities by assigning them the higher weights and make
recommendations on the basis of „value to voter‟ mechanism. By the use of these two
diverse applicable soft computing mechanisms, we have contributed following
achievement in the Chapter which is described as follows:
We have conferred a book recommendation method based on aggregation of
expert‟s decision. Thus the problem of book recommendation is converted into
decision making problem.
A fuzzy technique based method using Ordered Weighted Aggregation
(OWA) is employed for book recommendation. As far as our search is
concerned, we did not find any book recommendation approach employing
these techniques for the said techniques.
We have proposed an aggregation technique, Ordered Ranked Weighted
Aggregation (ORWA) which uses the rank of the rankers. The proposed
technique may be very useful for decision making problems where rankers
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
74 | P a g e
need to be taken into consideration. The ORWA is used for the
recommendation of books.
The rest of the Chapteris organized as follows; we have described ordered weighted
aggregation (OWA) and its applications are discussed. The importance of the
technique followed by procedure of book recommendation using OWA is also stated.
The recommendation technique based on ORWA and its advantages are discussed in
section 4.4 with suitable diagrams and examples. Section 4.5 deals with the
experimental results, in which there is a comprehensive discussion to highlight the
various aspects of results which is obtained by different experiments. Finally we have
summarized in section 3.6.
4.2 Ordered Weighted Aggregation
Ordered Weighted Aggregation (OWA) is a fuzzy based aggregation approach to
handle the uncertainty. Fuzzy techniques have been used widespread for various
scientific and daily life problems. Ordered Weighted Averaging (OWA) operator is a
well-known fuzzy based averaging operator which was introduced by R. Yager[247].
A variety of its applications have been presented in the literature. Several authors
have used OWA operator for various applications[248], [249]. The author [250] used
OWA operator based novel fuzzy queries for web searching. The researchers have
also applied the OWA operator‟s application in several GIS environments [251],
[252], [253].
Ordered weighted aggregation operator is very useful for aggregating multiple
criterions [235]. Mathematically we give OWA as;
1 2
1
, , ., = n
n k k
k
OWA x x x W Z
----------- (4.1)
Where Zk implies that if we re-order the values x1, x2 , … xn in descending order,
we get a sequence z1, z2, .. , zn i.e. z1≥ z2≥ ,… zn-1≥zn. The weights „Wk‟ for OWA
operator is calculated by using following equation [33, 25].
k / – 1 / ,W Q k m Q k m ----------- (4.2)
Where k = 1, 2… m.
Function Q(r) for relative quantifier can be calculated as:
0 if r<a
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
75 | P a g e
Q(r) =
if a≤ r ≤b --- (4.3)
1 if r>b
Where Q (0) = 0, ∃r ε [0, 1] such that Q(r) =1, and a, b and r ε [0,1]. By using
different linguistic quantifier for different a and b, we can find different weights.
E.g. „Most‟ is a linguistic quantifier for which a=0.3 and b=0.8, by the use of these
quantifiers those books are preferred which are recommended by most of the
universities. Thus each book is assigned a value, and upon these values the books
are sorted and ranked. Similarly, „as many as possible‟ and „at least half‟ are other
quantifiers for which values of (a, b) are (0.5, 1) and (0, 0.5) respectively. Graphical
representations of these fuzzy linguistic quantifiers are shown in the Figure 4.1,
Figure 4.2 and Figure 4.3 for „most‟, „as many as possible‟, and „at least half‟,
respectively.
Figure 4.1: Most Quantifier
Figure 4.2: As many as possible quantifier
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
76 | P a g e
Figure 4.3: At least half quantifier
4.3 Book Recommendation based on Ordered Weighted Aggregation
This section gives the description of OWA, its use and application. The book
recommendation approach using OWA is also discussed.Book recommendation
approach based on OWA is illustrated in this section. The example 3.1 is considered and
accordingly the weights are applied which are obtained as mentioned in example 4.1.
Example 4.1: For number of criteria (m) = 5 and parametric values as a=0 and b=0.5,
we will have corresponding weights for OWA values as:
w (1) =0.4, w (2) =0.4, w (3) =0.2, w (4) =0.0, w (5) =0.0.
In the same way, for a=0.3, b=0.8 the obtained of weights are; w (1) =0.0, w (2)
=0.2, w (3) =0.4, w (4) =0.4, w (5) =0.0.
For a=0.5, b=1.0 we have obtained values of weights as; w (1) =0.0, w (2) =0.0, w
(3) =0.2, w (4) =0.4, w (5) =0.4.
For obtaining the results of recommended top ranked books using relative quantifier
with OWA as an operator, we need to use above weights and employ it in example 3.1.
By using weights obtained in example 4.1, equation (4.1), and Table 3.5 of Chapter 3,
ranked books for these quantifiers can be obtained. The rankings are shown in Table
4.1, Table 4.2andTable 4.3, respectively.
Table 4.1:Ranked books using relative quantifier most
Rank Position Ranked Books
1 B2
2 B3
3 B1
4 B4
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
77 | P a g e
Table 4.2: Ranked books using relative quantifier As many as possible
Rank Position Ranked Books
1 B4
2 B2
3 B1
4 B3
Table 4.3: Ranked books using relative quantifier At least half
Rank Position Ranked Books
1 B2
2 B1
3 B3
4 B4
The basis of the whole recommendation process adopted is the top universities in
the QS ranking and their recommended books for enrolled student at respective
campuses. In PAS technique, we have not assigned any weights to the universities.
Hence, PAS can be perceived as un-weighted aggregation of scores assigned to
books using algorithm 3.1. Each course which consists of several books, are
assigned a score using PAS. By using OWA with different linguistic quantifiers, we
can assign weights to the university.
4.4 Book Recommendation based on Ordered Ranked Weighted
Aggregation (ORWA)
The modification in OWA for the situations where it is useful to incorporate value of
voters is discussed in the section and detail procedure of book recommendation is also
elaborated.
4.4.1 Ordered Ranked Weighted Aggregation (ORWA )
Ordered weighted aggregations (OWA) have been widely used in computational
intelligence because of its strength in modeling the multi criteria decision making
problems [253]. The OWA operator where weights assignment is guided by quantifier
[254] is found to be effective for the problems where criteria are well defined and the
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
78 | P a g e
voters need not be taken into consideration. However for a case where the value of
voters or rankers matter, i.e. a ranker A has some preference over ranker B then their
respective ranking should be assigned weights accordingly in the order, i.e. weights
assigned to A should be higher than the weights assigned to B. Considering the above
situation, we introduce a weight assignment formula for OWA and modify OWA
basic formula to obtain ordered ranked weighted aggregation (ORWA) operator that
takes into account the importance of the ranking agents.
The application of ORWA may influence several real life decision making
problems. In fact, ORWA may be ideally useful in all decision making problems
where the recommendations are given by the experts, and we need to weight the
experts or rankers. The ORWA would be very helpful in the recommendations of
voting results, sports team, universities preferences and web sites selections, etc.
We have proposed Ordered Ranked Weighted Aggregation operator to
recommend books in which the weight assignment procedure for OWA is modified so
that it may make use of the positional rank of the recommending agents. Keeping the
above concept in consideration, we have chosen top universities amongst the Indian
universities and their recommendation for different courses of computer science is
investigated, it is evident that a recommendation of a book by a high ranked
university will eventually increase the importance of the recommended books.
Aggregation weights „v‟ to different universities are assigned using formula;
vi=
----------- (4.4)
Where, n is the number of universities. N is given by; N = = .
And i=1,2,3…n. „i‟ indicates the ith
ranked university i.e. i=1 means first ranked
university, i=2 means second ranked university, and so on. Also the weights „vi‟ fulfill
the following conditions;
i. vi ε [0,1]
ii. i = 1.
Further, vi indicates the weights assigned to the ith
ranked university, i.e. the best
ranked university is associated with a maximum weight to it, and hence more
preferred over other least ranked universities. Thus vi>vj for i<j. i.e. for five different
ordered ranked universities U_1, U_2, U_3, U_4 and U_5, ordered in best to least
ranks; we have v1>v2>v3>v4>v5.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
79 | P a g e
4.4.2 Book Recommendation based on ORWA
Book recommendation approach based on ORWA is illustrated in this section. The
example 3.1 is considered and accordingly the weights are applied which are obtained
as mentioned in example 4.2.
Example 4.2: For five different universities we have m=5 that gives N = 15, thus we
get five values of weights as: v1 =1/3 =0.3333, v2 =4/15 =0.2666, v3 =1/5 =0.20, v4
=2/15 =0.1333, v5 =1/15 =0.0666.
We give formula to obtain ORWA as;
ORWA =1
n
i i
i
v y
----------- (4.5)
vi is given by the equation (4.4) and yi is the score given to a book by ith
ranked
university.
Referring to Table 5; we have preference scores of books for five ranked
universities. Considering equation 4.5, we get y1=1.0, y2=0.0, y3=0.66, y4=0.0 and
y5=0. 66.
ORWA = (0.3333 × 1) + (0.2666 × 0) + (0.20 × 0.66) + (0.1333 × 0) + (0.0666 ×
0.66)
= 0.0566
In the similar way we will be getting different values for book B2, B3 and B4. The
values obtained are as follows: B1 =0.4439292, B2= 0.619878, B3=0.309989,
B4=0.308556. The ranking of books illustrated using example 3.1 is tabulated in Table
4.4.
We can easily see the difference of the weights obtained by the OWA operator and
ORWA operator as calculated in section 4.2 and 4.3 respectively. The OWA operator
has W1 as 0 as well in several cases which would be associated with highest scored
ranker that eventually will make the final value zero. i.e. the most valuable ranker
may get „0‟ value whereas the ORWA operator, which has a modified way of
assigning weights to OWA, considers the strategy that highest weights should be
assigned to most valuable ranker, in our case the best ranked university.A specific
weight is associated to each university which is recommending a book, and ORWA
technique is used as described by equation (4.5). A block diagram for whole
procedure is given in
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
80 | P a g e
Figure 4.4. The detail discussion on the results is done in the section 4.5.
Figure 4.4: Ordered Ranked Weighted Aggregation based Book Recommendation System
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
81 | P a g e
Table 4.4: Ranked books based on Ordered Ranked Weighted Aggregation technique for example
3.1
Rank Position Ranked Books
1 B2
2 B1
3 B3
4 B4
4.5 Results and Discussions
In the previous sections of this chapter, we have described the procedure of
recommendation and generalized steps for the implementation of the recommendation
scheme is presented in respective sections. In this section, the results of the
recommendation by different techniques applied in the previous discussions, are
presented and discussed in details. The methods and steps of data collections, data
filtering, datasets, and pros and cons of the methods are also discussed.
4.5.1 Dataset
Basically, we are concerned in recommending books for university graduates of
Indian universities. Initially, those different books were taken that could be a part of
the curriculum of the universities. Though, there were neither any criteria nor any
limit of the inclusion of the books. Then we had a second thought to filter the data in a
way that could fulfil our objective, i.e. top ranked books of the course to the students.
For this, it seems adequate to include the top ranked institutions and their
recommended books. This step will filter the data as well as it helps in explaining why
the methodology (section 3.4) is chosen for the recommendation process. Also, only
„computer science‟ as a subject is selected from these universities/institutions.
Because, once we can find the method of presenting top books for any specific
subject, it can easily be extended for all other subjects. Hence, different courses of
computer science like, Discrete Mathematics, Data Structure, etc. which are almost
considered in top institutions, have been included. The methods of selection of
universities and courses, and the final recommendation by the different techniques are
discussed in subsequent sub sections.
4.5.2 Experimental Results
In this section, we have discussed the final recommendations by all 5 different
techniques implemented which include PAS, OWA with quantifiers „at least half‟, „as
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
82 | P a g e
many as possible‟, „most‟ and ORWA. We have applied all the techniques which are
discussed above. Different ranked books are obtained by using all these techniques.
The books are represented by unique course codes. E.g.: code „CD‟ has been used to
refer books on Compiler Design see table (3.14). In the similar way, different courses
of books have different notation for representations. For each book we have different
sequence of the books according to their ranking. For „Compiler Design‟ the different
codes are CD1, CD2, etc. The details of the books on compiler design including code,
author of the books, title and publisher, for which syllabus are available, are listed in
Table 4.5.
Table 4.5: Code and details for books on Compiler Design
course Code Author Title Publisher
CD.1. Alfred V. Aho, Monica
S. Lam, Ravi Sethi and
Jeffrey D. Ullman:
Compilers: Principles,
Techniques, and Tools
2/E, AddisonWesley
2007.
CD.2. Andrew Appel Modern Compiler
Implementation in
C/ML/Java
Cambridge University
Press, 2004
CD.3. Dick Grune, Henri E.
Bal, Cerial J.H.
Jacobs and Koen G.
Langendoen:
Modern Compiler
Design
John Wiley& Sons,
Inc. 2012.
CD.4. S. Muchnick, Advanced Compiler
Design &
Implementation,
Indian Reprint 2002.
CD.5. K Cooper,
L Torczon
Engineering a Compile
r
2nd Ed., Morgan Kauf
mann, 2011
CD.6. KC Louden, Compiler Construction:
Principles and Practice
Cengage Learning, 199
7
CD.7. D Grune et al. Modern
Compiler Design
Wiley, 2000
CD.8. Michael L Scott, Programming Languag
e Pragmatics
3rd Ed., Morgan Kauf
mann, 2009
CD.9. Tremblay, J.P. and
Sorenson, P.G.
Theory and Practice of
Compiler Writing
SR Publications.
2005
CD.10 Tremblay, J.P. and
Sorenson, P.G.
Parsing Techniques: A
Practical Guide
Ellis Horwood.
1998.
Course code of the books and corresponding rank given by respective universities
are given in Table 4.6.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
83 | P a g e
Table 4.6: Ranked list of book „compiler design‟ by top universities
Rank
Position U_1 U_2 U_3 U_4 U_5 U_6 U_7
11. CD1 CD2 CD1 CD1 CD1 - CD1
12. CD2 CD4 CD2 - CD4 - CD9
13. CD3 CD1 CD5 - - - CD5
14. - - CD6 - - - CD6
15. - - CD7 - - - CD10
16. - - CD8 - - - -
17. - - CD4 - - - -
18. - - - - - - -
19. - - - - - - -
20. - - - - - - -
From the table it is evident that U_1 has ranked book CD1 1st. CD2 and CD3 are
ranked 2nd
and 3rd
respectively. As U_1 is the 1st ranked university, this implies that
top university has recommended only three books on compiler design for their
students. The book CD1 is almost ranked by all the university except U_6, which has
not issued list of any book for the particular course. U_2 has ranked CD1 3rd
and CD2
is ranked 1st. However, U_3, U_4, U_5 and U_7 all have ranked CD1 1
st. It is also
observed that U_4 has recommended only one book. The PAS based ranking of books
are obtained by applying the procedure illustrated in example 3.1. To get the ranking
of books, say compiler design, we have considered all 10 books recommended by top
universities. These ranks are numerically represented in Table 4.10. In the table, R
(U_1) indicates rank given by U_1. These ranks are converted into scores and shown
in Table 3.14. As described in Table 3.2 of example 3.1in previous chapter, the best
rank i.e. „1‟ is assigned „-1‟, rank 2 is assigned „-2‟, and so on. The book which is not
ranked by any of the university is assigned lowest value which is „-8‟ here. Thus cells
values „0‟ of Table 3.13 have changed to „-8‟ in Table 3.14.
The final PAS is given in Table 3.15. In the section 3.3 and 3.4, the methods of
finding rank using OWA and ORWA are discussed. With the help of this procedure,
we can get the different scores for books and accordingly different rankings may be
obtained. The details of these scores, i.e. scores for books using PAS, OWA with
three differentquantifiers namely, as many as possible, most and at least half, and
ORWA are mentioned in Table 4.10.
The CD1 has almost attained maximum score from all the adopted methods. Its
score by PAS is 0.8252, by OWA with quantifiers at least half, as many as possible
and most are 0.8805, 0.9361 and 0.7142. However, CD1 has obtained 0.8285 by using
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
84 | P a g e
ORWA technique. For 2nd
position, CD2 has maximum value for three techniques
whereas other two methods have CD4 in second rank.
Table 4.7: Compiler design ranked books by top 7 Universities
Course
Code R (U_1) R (U_2) R (U_3) R (U_4) R (U_5) R (U_6) R (U_7)
CD.1. 1 3 1 1 1 0 1
CD.2. 2 1 2 0 0 0 0
CD.3. 3 0 0 0 0 0 0
CD.4. 0 2 7 0 2 0 0
CD.5. 0 0 3 0 0 0 3
CD.6. 0 0 4 0 0 0 4
CD.7. 0 0 5 0 0 0 0
CD.8. 0 0 6 0 0 0 0
CD.9. 0 0 0 0 0 0 2
CD.10. 0 0 0 0 0 0 5
Table 4.8: Rank to Score conversion of book Compiler Design
Course
Code S (U_1) S (U_2) S (U_3) S (U_4) S (U_5) S (U_6) S (U_7)
CD.1. -1 -3 -1 -1 -1 -8 -1
CD.2. -2 -1 -2 -8 -8 -8 -8
CD.3. -3 -8 -8 -8 -8 -8 -8
CD.4. -8 -2 -7 -8 -2 -8 -8
CD.5. -8 -8 -3 -8 -8 -8 -3
CD.6. -8 -8 -4 -8 -8 -8 -4
CD.7. -8 -8 -5 -8 -8 -8 -8
CD.8. -8 -8 -6 -8 -8 -8 -8
CD.9. -8 -8 -8 -8 -8 -8 -2
CD.10. -8 -8 -8 -8 -8 -8 -5
The score obtained using above techniques gives the corresponding ranking of the
books. Thus, we may have 5 different ranking of book, „compiler design‟. These
rankings with the method adopted are given in Table 3.16.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
85 | P a g e
Table 4.9: Positional Score for book Compiler Design
Course
Code
PS
(U_1)
PS
(U_2)
PS
(U_3) PS (U_4)
PS
(U_5)
PS
(U_6)
PS
(U_7) PAS
CD.1. 1 0.777 1 1 1 0 1 0.825
CD.2. 0.888 1 0.888 0 0 0 0 0.396
CD.3. 0.777 0 0 0 0 0 0 0.111
CD.4. 0 0.888 0.333 0 0.888 0 0 0.301
CD.5. 0 0 0.777 0 0 0 0.777 0.222
CD.6. 0 0 0.666 0 0 0 0.666 0.190
CD.7. 0 0 0.555 0 0 0 0 0.079
CD.8. 0 0 0.444 0 0 0 0 0.063
CD.9. 0 0 0 0 0 0 0.888 0.126
CD.10. 0 0 0 0 0 0 0.555 0.079
Table 4.10: Score obtained by recommendation approaches for compiler design
Course
Code PAS ORWA
OWA (At Least
half)
OWA (As many
as possible)
OWA
(most)
CD.1. 0.8252 0.8805 0.9361 0.7142 0.8285
CD.2. 0.3965 0.5947 0.7931 0 0.2283
CD.3. 0.111 0.1942 0.2219 0 0
CD.4. 0.3012 0.3447 0.3488 0.2537 0.3393
CD.5. 0.222 0.1664 0.2219 0.2219 0.1997
CD.6. 0.1902 0.1426 0.1902 0.1902 0.1712
CD.7. 0.0792 0.0990 0.1585 0 0.1426
CD.8. 0.0634 0.0792 0.1268 0 0.1141
CD.9. 0.1268 0.0317 0 0.2537 0
CD.10. 0.0792 0.0198 0 0.1585 0
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
86 | P a g e
Table 4.11: Five different ranking of book „compiler design‟
Rank
position
PAS based
ranking
ORWA
based
ranking
OWA (At
least) based
ranking
OWA (As
many as)
based
ranking
OWA (most)
based
ranking
1 CD.1. CD.1. CD.1. CD.1. CD.1.
2 CD.2. CD.2. CD.2. CD.4. CD.4.
3 CD.4. CD.4. CD.4. CD.9. CD.2.
4 CD.5. CD.3. CD.3. CD.5. CD.5.
5 CD.6. CD.5. CD.5. CD.6. CD.6.
6 CD.9. CD.6. CD.6. CD.10. CD.7.
7 CD.3. CD.7. CD.7. CD.2. CD.8.
8 CD.7. CD.8. CD.8. CD.3. CD.9.
9 CD.10. CD.9. CD.9. CD.7. CD.10.
10 CD.8. CD.10. CD.10. CD.8. CD.3.
Table 4.12: Five different ranking of book „Discrete Mathematics‟
Rank
Position
PAS based
ranking
ORWA
based
ranking
OWA (At
least half)
based
ranking
OWA (As
many as)
based
ranking
OWA (most)
based
ranking
1 DM.9. DM.1. DM.1. DM.9. DM.8.
2 DM.1. DM.8. DM.2. DM.10. DM.9.
3 DM.10. DM.9. DM.4. DM.12. DM.10.
4 DM.8. DM.2. DM.3. DM.16. DM.12.
5 DM.2. DM.3. DM.5. DM.8. DM.1.
6 DM.4. DM.10. DM.6. DM.15. DM.4.
7 DM.12. DM.4. DM.7. DM.17. DM.13.
8 DM.16. DM.5. DM.8. DM.13. DM.5.
9 DM.3. DM.6. DM.9. DM.18. DM.14.
10 DM.5. DM.7. DM.10. DM.14. DM.11.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
87 | P a g e
Table 4.13: Five different ranking of book ‟Artificial Intelligence‟
Rank Position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based ranking
OWA (As
many as)
based ranking
OWA (most)
based ranking
1 AI.2. AI.2. AI.2. AI.2. AI.2.
2 AI.5. AI.7. AI.7. AI.21. AI.14.
3 AI.21. AI.5. AI.5. AI.5. AI.20.
4 AI.7. AI.4. AI.4. AI.20. AI.15.
5 AI.11. AI.6. AI.6. AI.11. AI.21.
6 AI.4. AI.1. AI.1. AI.14. AI.16.
7 AI.6. AI.3. AI.3. AI.15. AI.7.
8 AI.1. AI.11. AI.8. AI.16. AI.17.
9 AI.14. AI.8. AI.9. AI.17. AI.8.
10 AI.15. AI.9. AI.10. AI.18. AI.18.
Table 4.14: Five different ranking of book „Data Structure‟
Rank Position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based ranking
OWA (As
many as)
based ranking
OWA (most)
based ranking
1 DS.1. DS.1. DS.1. DS.12.
DS.6.
2 DS.2. DS.2. DS.3. DS.6. DS.4.
3 DS.4. DS.4. DS.2. DS.4. DS.1.
4 DS.6. DS.6. DS.4. DS.2. DS.5.
5 DS.12.
DS.3. DS.5. DS.7. DS.7.
6 DS.3. DS.5. DS.6. DS.1. DS.8.
7 DS.7. DS.12.
DS.12.
DS.8. DS.9.
8 DS.5. DS.7. DS.7. DS.13.
DS.10.
9 DS.8. DS.8. DS.8. DS.9. DS.11.
10 DS.13.
DS.9. DS.9. DS.10.
DS.12.
Table 4.15: Five different ranking of book „Principal of Data Base‟
Rank Position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based
ranking
OWA (As
many as)
based
ranking
OWA (most)
based
ranking
1 DB.2. DB.2. DB.1. DB.2. DB.2.
2 DB.1. DB.1. DB.2. DB.4. DB.4.
3 DB.4. DB.4. DB.3. DB.1. DB.1.
4 DB.7. DB.3. DB.4. DB.7. DB.3.
5 DB.3. DB.7. DB.7. DB.5. DB.5.
6 DB.5. DB.5. DB.5. DB.6. DB.6.
7 DB.6. DB.6. DB.6. DB.8. DB.7.
8 DB.8. DB.8. DB.8. DB.9. DB.8.
9 DB.9. DB.9. DB.9. DB.13. DB.9.
10 DB.13. DB.10 DB.10 DB.10 DB.10
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
88 | P a g e
Table 4.16: Five different ranking of book „„Computer Graphics‟
Rank Position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based ranking
OWA (As
many as)
based ranking
OWA (most)
based ranking
1 CG.2. CG.2. CG.2. CG.14. CG.14.
2 CG.7. CG.7. CG.1. CG.16. CG.7
3 CG.1. CG.1. CG.3. CG.7. CG.15.
4 CG.3. CG.3. CG.4. CG.17. CG.16.
5 CG.14. CG.4. CG.5. CG.15. CG.17.
6 CG.16. CG.5. CG.6. CG.18. CG.18.
7 CG.4. CG.6. CG.7. CG.19. CG.19.
8 CG.17. CG.8. CG.8. CG.20. CG.20.
9 CG.5. CG.9. CG.9. CG.2. CG.2.
10 CG.15. CG.10. CG.10. CG.1. CG.1.
Table 4.17: Five different ranking of book „Software Engineering‟
Rank Position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based ranking
OWA (As
many as)
based ranking
OWA (most)
based ranking
1 SE.2. SE.2. SE.2. SE.2. SE.2.
2 SE.4. SE.3. SE.1. SE.18 SE.9.
3 SE.9. SE.1. SE.4. SE.9. SE.11
4 SE.3. SE.4. SE.5. SE.11 SE.12
5 SE.18 SE.9. SE.3. SE.4. SE.8.
6 SE.1. SE.5. SE.6. SE.12 SE.4.
7 SE.11 SE.6. SE.7. SE.15 SE.13
8 SE.5. SE.7. SE.8. SE.13 SE.5.
9 SE.8. SE.8. SE.9. SE.16 SE.14
10 SE.12 SE.10 SE.10 SE.14 SE.10
Table 4.18: Five different ranking of book „„Operating System‟
Rank Position Rank position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based ranking
OWA (As
many as)
based ranking
1 OS.2. OS.2. OS.2. OS.4. OS.4.
2 OS.4. OS.3. OS.3. OS.2. OS.2.
3 OS.3. OS.4. OS.1. OS.13. OS.3.
4 OS.5. OS.1. OS.4. OS.10. OS.5.
5 OS.13. OS.5. OS.5. OS.7. OS.7.
6 OS.1. OS.6. OS.6. OS.11. OS.8.
7 OS.10. OS.7. OS.7. OS.8. OS.9.
8 OS.7. OS.8. OS.8. OS.12. OS.6.
9 OS.11. OS.9. OS.9. OS.5. OS.10.
10 OS.8. OS.13. OS.13. OS.9. OS.11.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
89 | P a g e
Table 4.19: Five different ranking of book Computer Network‟
Rank Position PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based ranking
OWA (As
many as)
based ranking
OWA (most)
based ranking
1 CN.2. CN.2. CN.2. CN.2. CN.12.
2 CN.17. CN.1. CN.1. CN.17. CN.2.
3 CN.1. CN.3. CN.7. CN.12. CN.13.
4 CN.12. CN.12. CN.8. CN.13. CN.11.
5 CN.7. CN.4. CN.3. CN.14. CN.14.
6 CN.11. CN.5. CN.9. CN.18. CN.7.
7 CN.13. CN.6. CN.4. CN.1. CN.8.
8 CN.8. CN.7. CN.10. CN.15. CN.15.
9 CN.14. CN.8. CN.5. CN.19. CN.9.
10 CN.18. CN.9. CN.6. CN.16. CN.16.
Table 4.20: Five different ranking of book „Theory of Computation‟
Rank
Position
PAS based
ranking
ORWA based
ranking
OWA (At
least half)
based
ranking
OWA (As
many as)
based
ranking
OWA (most)
based
ranking
1 TOC.1. TOC.1. TOC.1. TOC.1. TOC.1.
2 TOC.2. TOC.2. TOC.2. TOC.6. TOC.4.
3 TOC.6. TOC.3. TOC.4. TOC.2. TOC.5.
4 TOC.4. TOC.4. TOC.3. TOC.7. TOC.6.
5 TOC.3. TOC.5. TOC.5. TOC.9. TOC.7.
6 TOC.5. TOC.6. TOC.6. TOC.8. TOC.8.
7 TOC.7. TOC.7. TOC.7. TOC.10. TOC.2.
8 TOC.9. TOC.8. TOC.8. TOC.11. TOC.9.
9 TOC.8. TOC.9. TOC.9. TOC.4. TOC.10.
10 TOC.10. TOC.10. TOC.10. TOC.3. TOC.11.
The first rank is achieved by CD1 from all the five different ways of recommendation.
However, five out of seven universities have ranked CD1 1st. This means 70% of
universities recommendation for the best book on the topic is same as the overall
recommendation of the proposed approaches. CD2 has obtained the 2nd
position in the
ranking by three rankings namely, PAS, ORWA and OWA with at least quantifier,
whereas rest of the two methods have recommended CD4 in 2nd
position. Although,
only three universities considers CD2 in their ranking, it has attained a good position
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
90 | P a g e
in the ranking as the all three universities involving CD2 in their recommendation
have high repute and acquire better position in ranking of universities than others.
Apart from CD1 and CD2 only CD4 is recommended by more than two universities in
their recommendations and hence its top position is obvious. All the three universities
which have considered CD2 in 2nd
position, recommends CD4 at third position. The
OWA with quantifier most has also ranked CD2 in third position as it has already
suggested CD4 to 2nd
position in its recommendations.
CD3 is recommended only by U_1 but has obtained rank 4 ORWA method
whereas others techniques have awarded a lower ranking to it except quantifier „at
least half‟. U_1 is the first ranked university and ORWA has a simple philosophy of
assigning a higher value to best voters, thus, inclusion of CD3 in higher positions
supports its approach. However, CD3 is ranked lower to CD5 by other techniques.
CD9 is ranked only by U_7 and no other universities have included it in their ranking,
still „as many as possible‟ prefers to rank CD9 on 3rd
position. Thus we can say from
the performance, ORWA and at least half is clearly recommending more appealing
and appropriate recommendations than other techniques. CD5 and CD6 both have two
number of inclusions in 7 ranking, and all have recommended CD5 above CD6, hence
is the final recommendation of the proposed approaches. CD7, CD8 and CD10 have
only one inclusion and are displaced to last two positions by both ORWA and „at least
half‟ quantifier. The last positions differ for other techniques. PAS and „as many as
possible‟ have recommended CD8 in last position and „most‟ quantifier has Cd3 in its
last position of ranking.
The above procedure is applied to all books which in turn give 5 different ranking
of each course of books. The five different ranking for all books are obtained in a
similar way the ranking of books on Compiler Design is achieved. These rankings are
presented from Table 4.11 to Table 4.20.
In these tables the ranked list of books for different approaches are listed. The
books are represented by their course code. These rankings can be very useful in
finding the top 10 or top 5 books on specified topic as well as it can give us the best
book, i.e. 1st ranked books of the course concerned. There are few courses whose 1
st
ranked book is common for all the methods as they all have recommended same book
in 1st position. Software Engineering, Compiler Design and Artificial Intelligence are
the courses for which the top universities have recommended same first ranked books.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
91 | P a g e
Though most of the top ranked books are recommended at top positions in majority,
i.e. 3 or 4 out of 5 approaches coincide in their first position ranking of books for
different courses. The consistency of first ranked positions may help readers to
comfortably choose the desired books. The tables clearly depicts that not all the top
10 ranked books are the same as the top universities‟ recommendation, i.e. if we sort
the books directly from university rank in a way that books by top university is ranked
on top, then it is not necessary that order of these books remain same in final
recommendation. However, as far as recommendation from ORWA is concerned, it
gives the ranking of books which is in most of the cases are directly proportional to
university ranks.
Also, the method of ORWA takes consideration of how many universities have
ranked the books? If a book is ranked first by 1st ranked university but no other
university has recommended the book; the books which are recommended by most of
the university would be preferred and ranked better. Consider the tables 3.21, ranked
list of books on „Artificial Intelligence‟ (AI) are listed. AI.1. is the book which is
ranked first by 1st ranked university, although it has not been recommended even in
top 3 positions by any of the methods. Instead, AI.2 is ranked top as it is
recommended by almost all the universities except U_6, hence has obtained 1st
position in the ranking.
4.6 Summary
We have introduced two different schemes to recommend books. First, a fuzzy based
aggregation scheme known as Ordered Weighted Aggregation (OWA) which has
been used in various domains but was never used in the recommendation of books to
the best of our knowledge is implemented for the recommendation of books.
Secondly, we have proposed an aggregation operator, „Ordered Ranked Weighted
Aggregation (ORWA)‟ and suggested a recommendation technique which exploits
proposed ORWA.
The Ordered Ranked Weighted Aggregation incorporates rank of the rankers to
emphasize the importance of the rankers. Because we believe, a book recommended
by best ranked institution must get high preference than a book which is
recommended by a lower ranked institution. The ORWA gives the ranking positions
of the recommended books, along with the total recommended books. The strength of
assigning weights to the rankers in the ORWA provides a better recommendation.
Chapter 4: Book Recommendation based on Soft Computing Approaches
___________________________________________________________________________________
92 | P a g e
Since we do not have any benchmark for ranking the books, we can rely on the best
ranked universities‟ recommendation (syllabus). The recommendation by the all
above techniques has been listed and presented. From the results of the
recommendation, it is obvious that the top ranked books have acquired different
positions in different ranking and have a slightly difference in first ranked books.
However, further discussion on the performance of the techniques regarding which
one is better will be discussed in Chapter 6.
We believe the proposed technique may meet the user‟s need and provide them the
perfect books they need. For the sake of illustration and ease of experiments, we have
shown the procedure considering books from top institutions. However, we can
generalize the procedure of the recommendation for any kind of items. The robustness
of the procedure may lead to a novel way in the field of recommendation and would
fulfill the demand of millions.
174 | P a g e
References:
[1] S. S. Sohail, J. Siddiqui, and R. Ali, “Book recommendation system using
opinion mining technique,” in International Conference on Advances in
Computing, Communications and Informatics (ICACCI), pp. 1609–1614,
2013.
[2] M. Hu, B. Liu, and S. M. Street, “Mining and Summarizing Customer
Reviews,” Proceedings of the tenth ACM SIGKDD international conference
on Knowledge discovery and data mining, USA. pp. 168-177, 2004.
[3] A. Andreevskaia and S. Bergler, “Mining WordNet for a Fuzzy Sentiment:
Sentiment Tag Extraction from WordNet Glosses.,” in EACL, vol. 6, pp. 209–
216, 2006.
[4] G. Carenini, R. T. Ng, and A. Pauls, “Interactive multimedia summaries of
evaluative text,” in Proceedings of the 11th international conference on
Intelligent user interfaces, pp. 124–131, 2006.
[5] S. S. Sohail, J. Siddiqui, and R. Ali, “Ordered Ranked Weighted Aggregation
based Book Recommendation Technique : A Link Mining Approach,” 14th
International Conference on Hybrid Intelligent Systems (HIS), IEEE, pp. 309–
314, 2014.
[6] H. K. Kim, H. Y. Oh, J. C. Gu, and J. K. Kim, “Commenders: A
recommendation procedure for online book communities,” Electron. Commer.
Res. Appl., vol. 10, no. 5, pp. 501–509, Sep. 2011.
[7] S.T. Yang and M.C. Hung, “A model for book inquiry history analysis and
book-acquisition recommendation of libraries,” Libr. Collect. Acquis. Tech.
Serv., vol. 36, no. 3–4, pp. 127–142, 2012.
[8] S. Pulakhandam and N. Patil, “Recommendation of Optimal Locations for
Government Funded Educational Institutes in Urban India Using a Hybrid
Data Mining Technique,” in Second International Conference on Advances in
Computing and Communication Engineering (ICACCE), pp. 560–567, 2015.
[9] J. Beel and S. Langer, “Research Paper Recommender Systems: A Literature
Survey,” Int. J. Digit. Libr., vol.17, no.4, 305-338, 2015.
References
175 | P a g e
[10] J. Beel, S. Langer, and M. Genzmehr, “Research paper recommender system
evaluation: A quantitative literature survey,” Proceedings of the Workshop on
Reproducibility and Replication in Recommender Systems Evaluation (RepSys)
at the ACM Recommender System Conference (RecSys), pp. 15–22, 2013.
[11] X. Wang and F. Yuan, “Course Recommendation by Improving BM25 to
Identity Students‟ Different Levels of Interests in Courses,” in International
Conference on New Trends in Information and Service Science, ( NISS), pp.
1372–1377, 2009.
[12] C.-Y. Huang, R.-C. Chen, and L.-S. Chen, “Course-recommendation system
based on ontology,” in 2013 International Conference on Machine Learning
and Cybernetics, vol. 3, pp. 1168-1173. IEEE, 2013.
[13] K. H. Tsai, T. C. Hsieh, T. K. Chiu, M. C. Lee, and T. I. Wang, “Automated
course composition and recommendation based on a learner intention,” in
Seventh IEEE International Conference on Advanced Learning Technologies
pp. 274–278, 2007.
[14] Ò. Celma, M. Ramírez, and P. Herrera, “Foafing the music: A music
recommendation system based on RSS feeds and user preferences,” In: 6th
International Conference on Music Information Retrieval (ISMIR), pp. 464–
467, 2005.
[15] M. T. Group and U. P. Fabra, “Foafing the Music : Bridging the semantic gap
in music recommendation,” Web Semantics: Science, Services and Agents on
the World Wide Web, vol. 6, no. 4, pp. 250-256, 2008:
[16] K. Yoshii, M. Goto, K. Komatani, T. Ogata, and H. G. Okuno, “Hybrid
Collaborative and Content-based Music Recommendation Using Probabilistic
Model with Latent User Preferences,” in ISMIR, vol. 6, pp. 296–301, 2006.
[17] F. Peleja and P. Dias, “A recommender system for the TV on the web :
integrating unrated reviews and movie ratings,” pp. 543–558, 2013.
[18] Z. Yu, X. Zhou, Y. Hao, and J. Gu, “TV program recommendation for
multiple viewers based on user profile merging,” User Model. User-adapt.
Interact., vol. 16, no. 1, pp. 63–82, 2006.
References
176 | P a g e
[19] J. M. Noguera, M. J. Barranco, R. J. Segura, and L. MartíNez, “A mobile 3D-
GIS hybrid recommender system for tourism,” Inf. Sci. (Ny)., vol. 215, pp. 37–
52, 2012.
[20] D. Gavalas, C. Konstantopoulos, K. Mastakas, and G. Pantziou, “Mobile
recommender systems in tourism,” J. Netw. Comput. Appl., vol. 39, pp. 319–
333, 2014.
[21] A. García-crespo, J. Chamizo, I. Rivera, M. Mencke, R. Colomo-palacios, and
J. M. Gómez-berbís, “SPETA : Social pervasive eTourism advisor,”
Telematics and Informatics, vol. 26, pp. 306–315, 2009.
[22] J. Shi, Y. Peng, and E. Erdem, “Simulation analysis on patient visit efficiency
of a typical VA primary care clinic with complex characteristics,” Simul.
Model. Pract. Theory, vol. 47, pp. 165–181, Sep. 2014.
[23] J.-H. Kim, J.-H. Lee, J.-S. Park, Y.-H. Lee, and K.-W. Rim, “Design of diet
recommendation system for healthcare service based on user information,” in
Fourth International Conference on Computer Sciences and Convergence
Information Technology, pp. 516–518, 2009.
[24] S. Saint-André, W. NeiraZalentein, D. Robin, and A. Lazartigues, “La
télépsychiatrie au service de l‟autisme,” Encephale., vol. 37, no. 1, pp. 18–24,
Feb. 2011.
[25] P. Pattaraintakorn, G. M. Zaverucha, and N. Cercone, “Web based health
recommender system using rough sets, survival analysis and rule-based expert
systems,” in International Workshop on Rough Sets, Fuzzy Sets, Data Mining,
and Granular-Soft Computing, pp. 491–499, 2007.
[26] C. Lee, M. Lee, D. Han, S. Jung, and J. Cho, “A framework for personalized
healthcare service recommendation,” in 10th International Conference on
ehealth Networking, Applications and Services, pp. 90–95, 2008.
[27] V. Kumar, K. M. P. D. Shrivastva, and S. Singh, “Cross Domain
Recommendation Using Semantic Similarity and Tensor Decomposition,”
Procedia Comput.Sci., vol. 85, pp. 317–324, 2016.
[28] S. S. Sohail, J. Siddiqui, and R. Ali, “Product Recommendation Techniques for
Ecommerce - past , present and future,” IJARCET, vol. 1, no. 9, pp. 219–225,
2012.
References
177 | P a g e
[29] I. Bose and R. K. Mahapatra, “Business data mining: a machine learning
perspective,” Information and Management, vol. 39, no.3, pp. 211–225, 2001.
[30] A. A. Shaikh and H. Karjaluoto, “Making the most of information technology
& systems usage: A literature review, framework and future research agenda,”
Comput. Human Behav., vol. 49, pp. 541–566, Aug. 2015.
[31] A. Karatzoglou, L. Baltrunas, and Y. Shi, “Learning to Rank for
Recommender Systems,” Proceedings of the 7th ACM conference on
Recommender systems. ACM, pp. 493–494, 2013.
[32] S. S. Sohail, J. Siddiqui, and R. Ali, “Recommender Systems for E-commerce :
In perspective of Business Strategies,” Shodh-Pioneer Journal of IT &
Management, vol.8, no.2, pp. 165-169, 2012.
[33] L. Getoor, “Link Mining : A New Data Mining Challenge,” ACM SIGKDD
Explorations Newsletter, vol. 5, no. 1, pp. 84–89, 2003.
[34] S. S. Sohail, J. Siddiqui, and R. Ali, “Book Recommendation Technique Using
Rank Based Scoring Method Abstract ,” National Conference on RIAIT, pp.
140-146, 2014.
[35] S. S. Sohail, J. Siddiqui, and R. Ali, “OWA based Book Recommendation
Technique,” Procedia Computer Science, vol. 62, pp. 126–133, 2015.
[36] S. S. Sohail, J. Siddiqui, and R. Ali, “Book Recommender System using Fuzzy
Linguistic Quantifier and Opinion Mining,” in The International Symposium
on Intelligent Systems Technologies and Applications, pp. 573–583, 2016.
[37] S. S. Sohail, J. Siddiqui, and R. Ali, “Feature extraction and analysis of online
reviews for the recommendation of books using opinion mining technique,”
Perspect. Sci., vol. 8, pp. 754–756, 2016.
[38] S. S. Sohail, J. Siddiqui, and R. Ali, “On Enhancement in Wearable Devices
using User Feed-back based Model,” International Journal of Computer
Science and Innovation, vol. 2016, no. 1, pp. 31–37, 2016.
[39] S. S. Sohail, J. Siddiqui, and R. Ali, “User feedback based evaluation of a
product recommendation system using rank aggregation method,” in Advances
in Intelligent Informatics, Springer, pp. 349–358, 2015.
References
178 | P a g e
[40] S. S. Sohail, J. Siddiqui, and R. Ali, “User Feedback Scoring and Evaluation of
a Product Recommendation System,” In Seventh International Conference on
Contemporary Computing (IC3), IEEE, pp. 525–530, 2014.
[41] D. Goldberg, D. Nichols, B. M. Oki, and D. Terry, “Using collaborative
filtering to weave an information tapestry,” Commun. ACM, vol. 35, no. 12,
pp. 61–70, 1992.
[42] G. Adomavicius and A. Tuzhilin, “Toward the next generation of
recommender systems: A survey of the state-of-the-art and possible
extensions,” IEEE Transactions on Knowledge and Data Engineering, vol. 17,
no. 6.pp. 734– 749, 2005.
[43] W. Hill, L. Stead, M. Rosenstein, and G. Furnas, “Recommending and
evaluating choices in a virtual community of use,” in Proceedings of the
SIGCHI conference on Human factors in computing systems, pp. 194–201,
1995.
[44] B. M. Sarwar, G. Karypis, J. a Konstan, and J. T. Riedl, “Application of
Dimensionality Reduction in Recommender System -- A Case Study,”
Architecture, vol. 1625, pp. 264–8, 2000.
[45] N. Good, J. Ben Schafer, J. A. Konstan, A. Borchers, B. Sarwar, J. Herlocker,
and J. Riedl, “Combining Collaborative Filtering with Personal Agents for
Better Recommendations,” In AAAI/IAAI, pp. 439-446, 1999.
[46] J.B. Schafer, J. A. Konstan, and J. Riedl. "E-commerce recommendation
applications." Applications of Data Mining to Electronic Commerce. Springer
US, pp. 115–153, 2001.
[47] R. Burke, “Hybrid Recommender Systems : Survey and,” User Model. User
Adapted Interact., vol. 12, no. 4, pp. 331–370, 2002.
[48] R. Burke, “Hybrid Web Recommender Systems,” LNCS 4321 - pp. 377–408,
2007.
[49] L. Candillier, F. Meyer, and M. Boullé, “Comparing state-of-the-art
collaborative filtering systems,” Lect. Notes Comput. Sci., vol. 4571, p. 548,
2007.
[50] X. Su and T. M. Khoshgoftaar, “A survey of collaborative filtering
techniques,” Adv. Artif. Intell., vol. 2009, no. Section 3, pp. 1–19, 2009.
References
179 | P a g e
[51] G. Shani and A. Gunawardana, “Evaluating recommendation systems,”
Recomm. Syst. Handb., pp. 257–298, 2011.
[52] D. H. Park, H. K. Kim, I. Y. Choi, and J. K. Kim, “Expert Systems with
Applications A literature review and classification of recommender systems
research,” Expert System with Applications., vol. 39, no. 11, pp. 10059–10072,
2012.
[53] X. Zhou, Y. Xu, Y. Li, A. Josang, and C. Cox, “The state-of-the-art in
personalized recommender systems for social networking,” Artificial
Intelligence Review vol.37, no. 2, pp. 119–132, 2012.
[54] J. Bobadilla, F. Ortega, A. Hernando, and A. Gutiérrez, “Recommender
systems survey,” Knowledge-based Syst., vol. 46, pp. 109–132, 2013.
[55] Y. Yang and X. Liu, “A re-examination of text categorization methods,” in
Proceedings of the 22nd annual international ACM SIGIR conference on
Research and development in information retrieval, pp. 42–49, 1999.
[56] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Analysis of recommendation
algorithms for e-commerce." In Proceedings of the 2nd ACM conference on
Electronic commerce, pp. 158-167, 2000.
[57] J. S. Breese, D. Heckerman, and C. Kadie, “Empirical analysis of predictive
algorithms for collaborative filtering,” in Proceedings of the Fourteenth
conference on Uncertainty in artificial intelligence, pp. 43–52, 1998.
[58] P. Resnick, P. Bergstrom, and J. Riedl, “GroupLens : An Open Architecture
for Collaborative Filtering of Netnews,” In Proceedings of the ACM
conference on Computer supported cooperative work, pp. 175–186, 1994.
[59] P. Resnick, H. R. Varian, and G. Editors, “Recommender Systems,”
Communications of the ACM, vol. 40, no. 3, pp. 56–58, 1997.
[60] W. Lin, “Efficient Adaptive-Support Association Rule Mining for
Recommender Systems,” Data mining and knowledge discovery vol.6, no.1
pp. 83–105, 2002.
[61] S. W. Ave, “Effective Personalization Based on Association Rule Discovery
from Web Usage Data,” In Proceedings of the 3rd international workshop on
Web information and data management, pp. 9–15, 2001.
References
180 | P a g e
[62] J. J. Sandvig, B. Mobasher, and R. Burke, “Robustness of Collaborative
Recommendation Based On Association Rule Mining,” In Proceedings of the
ACM conference on Recommender systems, pp. 105–111, 2007.
[63] M. O‟connor, D. Cosley, J. A. Konstan, and J. Riedl, “PolyLens: a
recommender system for groups of users,” in ECSCW, pp. 199–218, 2001.
[64] M. Anderson, M. Ball, H. Boley, S. Greene, N. Howse, D. Lemire, S.
McGrath, “RACOFI: A Rule-Applying Collaborative Filtering System,” Paper
presented at the conference IEEE/WIC COLA, Halifax, Canada. 13 October,
2003.
[65] K. Ali and W. Van Stam, “TiVo: making show recommendations using a
distributed collaborative filtering architecture,” in Proceedings of the tenth
ACM SIGKDD international conference on Knowledge discovery and data
mining, pp. 394–401, 2004.
[66] D. Lemire and S. Mcgrath, “Implementing a Rating-Based Item-to-Item
Recommender System in PHP / SQL,” D-01, On delette.com, pp. 84-98, 2005.
[67] Â. Cunningham and C. Hayes, “Smart radio - community based music radio,”
Knowledge-Based Systems vol.14, no. 3, pp.197-201, 2001.
[68] I. Barjasteh and R. Forsati, “Cold-Start Item and User Recommendation with
Decoupled Completion and Transduction,” In Proceedings of the 9th ACM
Conference on Recommender Systems, pp. 91-98, 2015.
[69] D. Library, I. Spectrum, M. Sites, C. Account, P. Sign, and R. Articles,
“Flycasting : using collaborative filtering to generate a playlist for online
radio,” In Proceedings of First International Conference on Web Delivering of
Music, pp. 123-130, 2001.
[70] M. P. Graus, “Improving the User Experience during Cold Start through
Choice-Based Preference Elicitation,” In Proceedings of the 9th ACM
Conference on Recommender Systems, pp. 273-276, 2015.
[71] M. J. Pazzani, “A framework for collaborative, content-based and
demographic filtering,” Artif. Intell. Rev., vol. 13, no. 5, pp. 393–408, 1999.
[72] G. Karypis, “Evaluation of Item-Based Top-N Recommendation Algorithms,”
Proc. tenth Int. Conf., no. TR 00--046, pp. 247–254, 2001.
References
181 | P a g e
[73] J. Gemmell, T. Schimoler, M. Ramezani, L. Christiansen, and B. Mobasher,
“Improving folkrank with item-based collaborative filtering,” Recomm. Syst.
Soc. Web, 2009.
[74] J. Gemmell, T. Schimoler, B. Mobasher, and R. Burke, “Recommendation by
Example in Social Annotation Systems,” In International Conference on
Electronic Commerce and Web Technologies, pp. 209–220, 2011.
[75] R. Jäschke, L. Marinho, A. Hotho, L. Schmidt-Thieme, and G. Stumme, “Tag
recommendations in folksonomies,” in European Conference on Principles of
Data Mining and Knowledge Discovery, pp. 506–514, 2007.
[76] A. Hotho, R. Jäschke, C. Schmitz, and G. Stumme, “Information retrieval in
folksonomies: Search and ranking,” in European Semantic Web conference,
pp. 411–426, 2006.
[77] N. Zheng and Q. Li, “A recommender system based on tag and time
information for social tagging systems,” Expert System with Applications., vol.
38, no. 4, pp. 4575–4587, 2011.
[78] J. Lu, D. Wu, M. Mao, W. Wang, and G. Zhang, “Recommender system
application developments : A survey,” Decision Support Systems, vol. 74, pp.
12–32, 2015.
[79] B. Marlin, “Modeling User Rating Profiles For Collaborative Filtering.” In
NIPS, pp. 627-634. 2003
[80] B. Marlin, “Collaborative filtering: A machine learning perspective.” Doctoral
dissertation, University of Toronto, 2004.
[81] J. W. Kim, B. H. Lee, M. J. Shaw, H.-L. Chang, and M. Nelson, “Application
of decision-tree induction techniques to personalized advertisements on
internet storefronts,” Int. J. Electron. Commer., vol. 5, no. 3, pp. 45–62, 2001.
[82] B. N. Miller, J. A. Konstan, and J. Riedl, “PocketLens: Toward a personal
recommender system,” ACM Trans. Inf. Syst., vol. 22, no. 3, pp. 437–476,
2004.
[83] R. Garfinkel, R. Gopal, A. Tripathi, and F. Yin, “Design of a shopbot and
recommender system for bundle purchases,” Decision Support Systems, vol.
42, pp. 1974–1986, 2006.
References
182 | P a g e
[84] X. Su and T. M. Khoshgoftaar, “Collaborative filtering for multi-class data
using belief nets algorithms,” in 18th IEEE International Conference on Tools
with Artificial Intelligence, pp. 497–504, 2006.
[85] R. M. Bell, Y. Koren, P. Ave, and F. Park, “Improved Neighborhood-based
Collaborative Filtering,” pp. 7–14, 2007.
[86] N. Sahoo, V. P. Singh, and T. Mukhopadhyay, "A hidden Markov model for
collaborative filtering," MIS Quarterly, vol. 6, no. 4, pp. 1329–1356, 2012.
[87] K. Verstrepen, “Top-N Recommendation for Shared Accounts,” In
Proceedings of the 9th ACM Conference on Recommender Systems pp. 59–
66, 2015.
[88] M. Aharon, D. Drachsler-cohen, O. Anava, N. Avigdor-elgrabli, and O.
Somekh, “ExcUseMe : Asking Users to Help in Item Cold-Start
Recommendations,” In Proceedings of the 9th ACM Conference on
Recommender Systems, pp. 83–90, 2015.
[89] R. R. Yager, “Fuzzy logic methods in recommender systems,” Fuzzy Sets
Syst., vol. 136, no. 2, pp. 133–149, 2003.
[90] S. Russell, P. Norvig, and A. Intelligence, “A modern approach,” Artif. Intell.
Prentice-Hall, Egnlewood Cliffs, vol. 25, p. 27, 1995.
[91] S. Weng and M. Liu, “Feature-based recommendations for one-to-one
marketing,” Expert Systems with Applications, vol. 26, pp. 493–508, 2004.
[92] K. W. Cheung, K. C. Tsui, and J. Liu, “Extended latent class models for
collaborative recommendation,” IEEE Trans. Syst. Man, Cybern. A Syst.
Humans, vol. 34, no. 1, pp. 143–148, 2004.
[93] R. S. Chen, Y. S. Tsai, K. C. Yeh, D. H. Yu, and Y. Bak-Sau, “Using data
mining to provide recommendation service,” WSEAS Trans. Inf. Sci. Appl.,
vol. 5, no. 4, pp. 459–474, 2008.
[94] A. S. Tewari and K. Priyanka, “Book recommendation system based on
collaborative filtering and association rule mining for college students,” in
International Conference on Contemporary Computing and Informatics
(IC3I), pp. 135–138, 2014.
[95] M. Mikawa, S. Izumi, and K. Tanaka, “Book recommendation signage system
using silhouette-based gait classification,” in 10th International Conference on
References
183 | P a g e
Machine Learning and Applications and Workshops (ICMLA), vol. 1, pp. 416–
419, 2011.
[96] J. Salter and N. Antonopoulos, “CinemaScreen recommender agent:
combining collaborative and content-based filtering,” IEEE Intell. Syst., vol.
21, no. 1, pp. 35–41, 2006.
[97] G. Kazai, D. Clarke, and M. Venanzi, “A Personalised Reader for Crowd
Curated Content,” In Proceedings of the 9th ACM Conference on
Recommender Systems, pp. 325–326, 2015.
[98] K. Lang, “Newsweeder: Learning to filter netnews,” in Proceedings of the
12th international conference on machine learning, pp. 331–339, 1995.
[99] R. J. M. and L. Roy, “Content-based book recommendation using learning for
text categorization,” In Proceedings of the 5th ACM Conference on Digital
Libraries, pp. 195-204, 2000.
[100] P. Jomsri, “Book recommendation system for digital library based on user
profiles by using association rule,” in Fourth International Conference on
Innovative Computing Technology (INTECH), pp. 130–134, 2014.
[101] D. Billsus, M. J. Pazzani, and J. Chen, “A Learning Agent for Wireless News
Access,” In Proceedings of the 5th international conference on Intelligent user
interfaces, pp. 33–36, 2000.
[102] Y. Blanco-fernández, J. J. Pazos-arias, A. Gil-solla, and M. Ramos-cabrer,
“Providing Entertainment by Content-based Filtering and Semantic Reasoning
in Intelligent Recommender Systems,” IEEE Transactions on Consumer
Electronics, vol.54, no.2, May 2008.
[103] T. Bansal, “Content Driven User Profiling for
Comment-Worthy Recommendations of News and Blog Articles Categories
and Subject Descriptors,” RecSys‟15, Vienna, Austria, pp. 195–202, 2015.
[104] M. de Gemmis, P. Lops, C. Musto, F. Narducci, and G. Semeraro,
“Semanticsaware content-based recommender systems,” in Recommender
Systems Handbook, Springer, pp. 119–159, 2015.
[105] J. Kyeong, Y. Ho, W. Ju, J. Ran, and J. Hae, “A personalized recommendation
procedure for Internet shopping support,” Electronic Commerce Research and
Applications, vol. 1, pp. 301–313, 2002.
References
184 | P a g e
[106] Z. Zeng, “An Intelligent E-commerce Recommender System Based on Web
Mining,” International Journal of Business and Management, vol. 4, no. 7, pp.
10–14, 2009.
[107] D. R. Liu and Y. Y. Shih, “Integrating AHP and data mining for product
recommendation based on customer lifetime value,” Inf. Manag., vol. 42, no.
3, pp. 387–400, 2005.
[108] P. Melville, R. J. Mooney, and R. Nagarajan, “Content-Boosted Collaborative
Filtering for Improved Recommendations,” In Proceedings of the Eighteenth
National Conference on Artificial Intelligence(AAAI-2002), pp. 187-192,
Edmonton, Canada, July, pp. 187–192, 2002.
[109] B. Krulwich, “Lifestyle finder: Intelligent user profiling using large-scale
demographic data,” AI Mag., vol. 18, no. 2, p. 37, 1997.
[110] L. Safoury and A. Salah, “Exploiting User Demographic Attributes for Solving
Cold-Start Problem in Recommender System,” Lecture Notes on Software
Engineering, vol. 1, no. 3, pp. 1–5, August 2013.
[111] B. Towle and C. Quinn, “Knowledge Based Recommender Systems Using
Explicit user models,” In Proceedings of the AAAI Workshop on
KnowledgeBased Electronic Markets pp. 74–77, 2000.
[112] E. Vildjiounaite, V. Kyllönen, T. Hannula, and P. Alahuhta, “Unobtrusive
dynamic modelling of tvprogramme preferences in a finnish household,”
Multimed. Syst., vol. 15, no. 3, pp. 143–157, 2009.
[113] R. D. Burke, K. J. Hammond, and B. C. Yound, “The FindMe approach to
assisted browsing,” IEEE Expert, vol. 12, no. 4, pp. 32–40, 1997.
[114] R. D. Burke, K. J. Hammond, and B. C. Young, “Knowledge-based navigation
of complex information spaces,” in Proceedings of the national conference on
artificial intelligence, vol. 462, p. 468, 1996.
[115] J. F. McCarthy, “Pocket restaurantfinder: A situated recommender system for
groups,” in Workshop on Mobile Ad-Hoc Communication at the 2002 ACM
Conference on Human Factors in Computer Systems, 2002.
[116] M. Biňas and E. Pietriková, “Useful recommendations for successful
implementation of programming courses,” in IEEE 12th International
References
185 | P a g e
Conference on Emerging eLearning Technologies and Applications (ICETA),
pp. 397–401, 2014.
[117] Y. Jing, “A Study of the Mass Customization-Based Strategy for the
Recommendation of Online Course Resources of the Open University of
China,” in 8th International Symposium on Computational Intelligence and
Design (ISCID), vol. 2, pp. 311–314, 2015.
[118] G. Koutrika, B. Bercovitz, R. Ikeda, F. Kaliszan, H. Liou, and H. Garcia-
Molina, “Flexible recommendations for course planning,” 2008.
[119] S. B. Aher and L. M. R. J. Lobo, “Combination of machine learning
algorithms for recommendation of courses in E-Learning System based on
historical data,” Knowledge-Based Systems, vol. 51, pp. 1–14, 2013.
[120] R. Burke, “Knowledge-based recommender systems,” Encycl. Libr. Inf. Syst.,
vol. 69, no. Supplement 32, pp. 175–186, 2000.
[121] A. Felfernig and R. Burke, “Constraint-based recommender systems:
technologies and research issues,” in Proceedings of the 10th international
conference on Electronic commerce, p. 3, 2008.
[122] D. Bridge, M. H. Göker, L. McGinty, and B. Smyth, “Case-based
recommender systems,” Knowl. Eng. Rev., vol. 20, no. 3, pp. 315–320, 2005.
[123] R. Burke, “Integrating knowledge-based and collaborative-filtering
recommender systems,” in Proceedings of the Workshop on AI and Electronic
Commerce, pp. 69–72, 1999.
[124] A. Felfernig and R. Burke, “Constraint-based Recommender Systems :
Technologies and Research Issues,” In Proceedings of the 10th international
conference on Electronic commerce, pp. 3, 2008.
[125] B. Smyth, “Case-Based Recommendation,” The Adaptive Web, LNCS 4321,
pp. 342–376, 2007.
[126] D. McSherry, “Minimizing dialog length in interactive case-based reasoning,”
in Proceedings of the 17th international joint conference on Artificial
intelligence-Volume 2, pp. 993–998, 2001.
[127] D. McSherry, “Increasing dialogue efficiency in case-based reasoning without
loss of solution quality,” in IJCAI, pp. 121–126, 2003.
References
186 | P a g e
[128] K. McCarthy, J. Reilly, L. McGinty, and B. Smyth, “Thinking
positivelyexplanatory feedback for conversational recommender systems,” in
Proceedings of the European Conference on Case-Based Reasoning
(ECCBR‟04) Explanation Workshop, pp. 115–124, 2004.
[129] A. Felfernig, G. Friedrich, K. Isak, K. Shchekotykhin, E. Teppan, and
D. Jannach, “Automated debugging of recommender user interface
descriptions,” Appl. Intell., vol. 31, no. 1, pp. 1–14, 2009.
[130] A. Felfernig, K. Isak, K. Szabo, and P. Zachar, “The VITA Financial Services
Sales Support Environment,” Association for the Advancement of Artificial
Intelligence, pp. 1692–1699, 2007.
[131] A. Felfernig, “Standardized configuration knowledge representations as
technological foundation for mass customization,” IEEE Trans. Eng. Manag.,
vol. 54, no. 1, pp. 41–56, 2007.
[132] Z. Jiang, W. Wang, and I. Benbasat, “Multimedia-based interactive advising
technology for online consumer decision support,” Commun. ACM, vol. 48,
no. 9, pp. 92–98, 2005.
[133] P. Auteri and R. Turrin, "Personalized Catch-up & DVR: VOD or Linear, That
is the Question," In Proceedings of the 9th ACM Conference on Recommender
Systems, pp. 227-227, 2015.
[134] M. Balabanović and Y. Shoham, “Fab: content-based, collaborative
recommendation,” Commun. ACM, vol. 40, no. 3, pp. 66–72, 1997.
[135] Q. Li and B. M. Kim, “Clustering approach for hybrid recommender system,”
in Proceedings of IEEE/WIC International Conference on Web Intelligence,
pp. 33–38, 2003.
[136] R. Hu and P. Pu, “Using personality information in collaborative filtering for
new users,” Recomm. Syst. Soc. Web, vol. 17, 2010.
[137] I. Soboroff and C. Nicholas, “Combining content and collaboration in text
filtering,” in Proceedings of the IJCAI, vol. 99, pp. 86–91, 1999.
[138] B. Smyth and P. Cotter, “A personalized television listings service,” Commun.
ACM, vol. 43, no. 8, pp. 107–111, 2000.
[139] A. Ansari, S. Essegaier, and R. Kohli, “Internet recommendation systems,”
American Marketing Association, pp. 363-375, 2000.
References
187 | P a g e
[140] F. Ortega, J. Sánchez, J. Bobadilla, and A. Gutiérrez, “Improving collaborative
filtering-based recommender systems results using Pareto dominance,”
Information Sciences, vol. 239, pp. 50–61, 2013.
[141] C. Basu, H. Hirsh, and W. Cohen, “Recommendation as classification: Using
social and content-based information in recommendation,” In Technical
Report, AAAI Press, Menlo Park, pp. 714-720, 1998.
[142] S. E. Middleton, N. R. Shadbolt, and D. C. De Roure, “Ontological
user profiling in recommender systems,” ACM Trans. Inf. Syst., vol. 22, no. 1,
pp. 54–88, 2004.
[143] A. Popescul, D. M. Pennock, and S. Lawrence, “Probabilistic models for
unified collaborative and content-based recommendation in sparse-data
environments,” in Proceedings of the Seventeenth conference on Uncertainty
in artificial intelligence, pp. 437–444, 2001.
[144] T. Hofmann and J. Puzicha, “Latent class models for collaborative filtering,”
in IJCAI, 1999, vol. 99, no. 1999.
[145] M. Claypool, A. Gokhale, T. Miranda, P. Murnikov, D. Netes, and M. Sartin,
“Combining content-based and collaborative filters in an online newspaper,” in
Proceedings of ACM SIGIR workshop on recommender systems, vol. 60, 1999.
[146] J. K. Kim, H. K. Kim, H. Y. Oh, and Y. U. Ryu, “A group recommendation
system for online communities,” Int. J. Inf. Manage., vol. 30, no. 3, pp. 212–
219, 2010.
[147] M. Brown-sica, “Using academic courses to generate data for use in evidence
based library planning,” The Journal of Academic Librarianship, vol. 39, no.3,
pp.275-287, 2013.
[148] D. Billsus and M. J. Pazzani, “User modeling for adaptive news access,” User
Model. User-adapt. Interact., vol. 10, no. 2–3, pp. 147–180, 2000.
[149] A. Sami, R. Nagatomi, M. Terabe, and K. Hashimoto, “Design of physical
activity recommendation system,” in Multi Conference on Informatics and
Computer Science and Information Systems, (MCCSIS-IADIS) Data Mining,
pp. 148-152, 2008.
References
188 | P a g e
[150] M. Wiesner and D. Pfeifer, “Adapting recommender systems to the
requirements of personal health record systems,” in Proceedings of the 1st
ACM International Health Informatics Symposium, pp. 410–414, 2010.
[151] L. Fernandez-Luque, R. Karlsen, and L. K. Vognild, “Challenges and
opportunities of using recommender systems for personalized health
education.,” in MIE, pp. 903–907, 2009.
[152] L. C. Braat and R. de Groot, “The ecosystem services agenda:bridging the
worlds of natural science and economics, conservation and development, and
public and private policy,” Ecosyst. Serv., vol. 1, no. 1, pp. 4–15, Jul. 2012.
[153] I. Gonzalez-Carrasco, R. Colomo-Palacios, J. L. Lopez-Cuadrado, Á. Garcı,
and B. Ruiz-Mezcua, “PB-ADVISOR: A private banking multi-investment
portfolio advisor,” Inf. Sci. (Ny)., vol. 206, pp. 63–82, 2012.
[154] M. Tang, Y. Jiang, J. Liu, and X. Liu, “Location-aware collaborative filtering
for QoS-based service recommendation,” in 19th IEEE International
Conference on Web Services (ICWS), pp. 202–209, 2012.
[155] Y. Koren, “Collaborative filtering with temporal dynamics,” Commun. ACM,
vol. 53, no. 4, pp. 89–97, 2010.
[156] J. Gemmell, T. Schimoler, M. Ramezani, L. Christiansen, and B. Mobasher,
“Improving FolkRank With Item-Based Collaborative Filtering,”
Recommender Systems & the Social Web, Oct 25 2009.
[157] D. Godoy and A. Amandi, “Hybrid content and tag-based profiles for
recommendation in collaborative tagging systems,” in Latin American Web
Conference, 2008. LA-WEB‟08., pp. 58–65, 2008,.
[158] L. Iaquinta, M. De Gemmis, P. Lops, G. Semeraro, M. Filannino, and P.
Molino, “Introducing serendipity in a content-based recommender system,” in
Eighth International Conference on Hybrid Intelligent Systems (HIS‟08), pp.
168–173, 2008.
[159] A. Nanopoulos, D. Rafailidis, P. Symeonidis, and Y. Manolopoulos,
“Musicbox: Personalized music recommendation based on cubic analysis of
social tags,” IEEE Trans. Audio. Speech. Lang. Processing, vol. 18, no. 2, pp.
407–412, 2010.
References
189 | P a g e
[160] H. W. Tung and V. W. Soo, “A personalized restaurant recommender agent for
mobile e-service,” in IEEE International Conference on e-Technology,
eCommerce and e-Service (EEE‟04), pp. 259–262, 2004.
[161] W. Woerndl, C. Schueller, and R. Wojtech, “A hybrid recommender system
for context-aware recommendations of mobile applications,” in 23rd
International Conference on Data Engineering Workshop, pp. 871–878, 2007.
[162] M. Brunato and R. Battiti, “PILGRIM: A location broker and mobility-aware
recommendation system,” in Proceedings of the First IEEE International
Conference on Pervasive Computing and Communications, pp. 265–272,
2003.
[163] M. Sarwat, J. J. Levandoski, A. Eldawy, and M. F. Mokbel, “LARS*:
An efficient and scalable location-aware recommender system,” IEEE Trans.
Knowl.Data Eng., vol. 26, no. 6, pp. 1384–1399, 2014.
[164] J. J. Levandoski, M. Sarwat, A. Eldawy, and M. F. Mokbel, “Lars: A
locationaware recommender system,” in IEEE 28th International Conference
on Data Engineering (ICDE), pp. 450–461, 2012.
[165] W. S. Yang, H. C. Cheng, and J. B. Dia, “A location-aware recommender
system for mobile shopping environments,” Expert System with Applications.,
vol. 34, no. 1, pp. 437–445, 2008.
[166] M. H. Park, J. H. Hong, and S. B. Cho, “Location-based recommendation
system using Bayesian user‟s preference model in mobile devices,” in
International Conference on Ubiquitous Intelligence and Computing, pp.
1130–1139, 2007.
[167] K. Božidar, O. Dijana, and B. Nina, “Temporal recommender systems,” in
Proceedings of the 10th WSEAS international conference on Applied computer
and applied computational science, pp. 248–253, 2011.
[168] T. Q. Lee, Y. Park, and Y.-T. Park, “A time-based approach to effective
recommender systems using implicit feedback,” Expert System with
Applications., vol. 34, no. 4, pp. 3055–3062, 2008.
[169] N. Lathia, S. Hailes, L. Capra, and X. Amatriain, “Temporal diversity in
recommender systems,” in Proceedings of the 33rd international ACM SIGIR
References
190 | P a g e
conference on Research and development in information retrieval, pp. 210–
217, 2010.
[170] F. Ullah, G. Sarwar, S. C. Lee, Y. K. Park, K. D. Moon, and J. T. Kim,
“Hybrid recommender system with temporal information,” in International
Conference onInformation Networking (ICOIN), pp. 421–425, 2012.
[171] Ò. Celma, “Foafing the music: Bridging the semantic gap in music
recommendation,” in International Semantic Web Conference, pp. 927–934,
2006.
[172] J. He and W. W. Chu, “A social network-based recommender system (SNRS),”
Physics Reports, Springer, vol.519, no.1, pp.1-49, 2010.
[173] J.-C. Wang and C.-C. Chiu, “Recommending trusted online auction sellers
using social network analysis,” Expert System with Applications., vol. 34, no.
3, pp. 1666– 1679, 2008.
[174] C. Cornelis, X. Guo, J. Lu, and G. Zhang, “A Fuzzy Relational Approach to
Event Recommendation,” 2nd Indian International Conference on Artificial
Intelligence (IICAI), pp. 2231–2242, 2007.
[175] C. Cornelis, J. Lu, X. Guo, and G. Zhang, “One-and-only item
recommendation with fuzzy logic techniques,” Information Sciences, vol. 177,
pp. 4906–4921, 2007.
[176] Z. Zhang, H. Lin, K. Liu, D. Wu, G. Zhang, and J. Lu, “A hybrid fuzzy-based
personalized recommender system for telecom products/services,” Inf. Sci.
(Ny)., vol. 235, pp. 117–129, 2013.
[177] M. Y. H. Al-shamri and K. K. Bharadwaj, “Fuzzy-genetic approach to
recommender systems based on a novel hybrid user model,” Expert systems
with applications, vol. 35, pp. 1386– 1399, 2008.
[178] M. J. Pazzani, “A framework for collaborative, content-based and
demographic filtering,” Artif. Intell. Rev., vol. 13, no. 5–6, pp. 393–408, 1999.
[179] M. Deshpande and G. Karypis, “Item-Based Top- N Recommendation
Algorithms,” vol. 22, no. 1, pp. 143–177, 2004.
[180] G. Karypis, “Evaluation of item-based top-n recommendation algorithms,” in
Proceedings of the tenth international conference on Information and
knowledge management, pp. 247–254, 2001.
References
191 | P a g e
[181] G. Shani and A. Gunawardana, “Evaluating Recommendation Systems,” In
Recommender systems handbook, pp. 257-297, 2011.
[182] W. Lin, S. A. Alvarez, and C. Ruiz, "Collaborative recommendation via
adaptive association rule mining," Data Mining and Knowledge Discovery,
vol.6, pp.83-105, 2000.
[183] P. Symeonidis, A. Nanopoulos, A. N. Papadopoulos, and Y. Manolopoulos,
“Collaborative recommender systems: Combining effectiveness and
efficiency,” Expert System with Applications, vol. 34, no. 4, pp. 2995–3013,
May 2008.
[184] S. Gong, “A collaborative filtering recommendation algorithm based on user
clustering and item clustering,” J. Softw., vol. 5, no. 7, pp. 745–752, 2010.
[185] B Sarwar, G Karypis, J Konstan, J Riedl, “Item-Based Collaborative Filtering
Recommendation Algorithms,” In Proceedings of the 10th international
conference on World Wide Web, pp. 285-295, 2001.
[186] J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl, “Evaluating
Collaborative Filtering Recommender Systems,” vol. 22, no. 1, pp. 5–53,
2004.
[187] R. Garcia and X. Amatriain, “Weighted content based methods for
recommending connections in online social networks,” in Workshop on
Recommender Systems and the Social Web, pp. 68–71, 2010.
[188] M. J. Pazzani and D. Billsus, “Content-Based Recommendation Systems,”
Adapt. Web, vol. 4321, pp. 325–341, 2007.
[189] T. Joachims, “Text categorization with support vector machines: Learning
with many relevant features,” in European conference on machine learning,
pp. 137–142, 1998.
[190] W. Huang, Y. Zhao, S. Yang, and Y. Lu, “Analysis of the user behavior and
opinion classification based on the BBS,” Appl. Math. Comput., vol. 205, no.
2, pp. 668–676, 2008.
[191] P. Lops, M. De Gemmis, and G. Semeraro, “Content-based Recommender
Systems : State of the Art and Trends,” In Recommender systems handbook,
pp. 73-105, Springer, US, 2011.
References
192 | P a g e
[192] R. J. Mooney and L. Roy, “Content-based book recommending using learning
for text categorization,” in Proceedings of the fifth ACM conference on Digital
libraries, pp. 195–204, 2000.
[193] C. Romero and S. Ventura, “Educational data mining: A survey from 1995 to
2005,” Expert System with Applications., vol. 33, no. 1, pp. 135–146, 2007.
[194] P. Taylor and R. R. Yager, “Multicriteria decision-making using fuzzy
measures," Cybernetics and Systems, vol. 46, no. 3-4, pp.150-171, August
2015.
[195] D. Anand and B. S. Mampilli, “Folksonomy-based fuzzy user profiling for
improved recommendations,” Expert System with Applications., vol. 41, no. 5,
pp. 2424–2436, Apr. 2014.
[196] E. S. Han and G. Karypis, “Feature-Based Recommendation System,” In
Proceedings of the 14th ACM international conference on Information and
knowledge management, pp. 446-452, 2005.
[197] K. Choi, D. Yoo, G. Kim, and Y. Suh, “A hybrid online-product
recommendation system: Combining implicit rating-based collaborative
filtering and sequential pattern analysis,” Electron. Commer. Res. Appl., vol.
11, no. 4, pp. 309–317, 2012.
[198] N. Pellas and I. Kazanidis, “Online and hybrid university-level courses with
the utilization of Second Life : Investigating the factors that predict student
choice in Second Life supported online and hybrid university-level courses,”
Computers in Human Behavior, vol. 40, pp. 31–43, 2014.
[199] P. C. Vaz, D. M. De Matos, B. Martins, and P. Calado, “Improving a Hybrid
Literary Book Recommendation System through Author Ranking Categories
and Subject Descriptors,” In Proceedings of the 12th ACM/IEEE-CS joint
conference on Digital, pp. 387–388, 2012.
[200] S. Maneeroj and A. Takasu, “Hybrid recommender system using latent
features,” in International Conference on Advanced Information Networking
and Applications Workshops WAINA‟09., pp. 661–666, 2009.
[201] S. K. Shinde and U. Kulkarni, “Hybrid personalized recommender system
using centering-bunching based clustering algorithm,” Expert System with
Applications., vol. 39, no. 1, pp. 1381–1387, 2012.
References
193 | P a g e
[202] T. Tran and R. Cohen, “Hybrid recommender systems for electronic
commerce,” in Proc. Knowledge-Based Electronic Markets, Papers from the
AAAI Workshop, Technical Report WS-00-04, AAAI Press, 2000.
[203] R. E. Krainer, “Towards a program for financial stability,” J. Econ. Behav.
Organ., vol. 85, pp. 207–218, Jan. 2013.
[204] L. M. De Campos, J. M. Fernández-Luna, J. F. Huete, and M. A.
RuedaMorales, “Combining content-based and collaborative
recommendations: A hybrid approach based on Bayesian networks,” Int. J.
Approx. Reason., vol. 51, no. 7, pp. 785–799, 2010.
[205] R. Burke, “Hybrid Web Recommender Systems,” The Adaptive Web, LNCS
4321 pp. 377–408, 2007.
[206] O. Kaššák, M. Kompan, and M. Bieliková, “Personalized hybrid
recommendation for group of users: Top-N multimedia recommender,” Inf.
Process.Manag., vol. 52, no. 3, pp. 459–477, 2016.
[207] G. Badaro, H. Hajj, W. El-Hajj, and L. Nachman, “A hybrid approach with
collaborative filtering for recommender systems,” in 9th International
Wireless Communications and Mobile Computing Conference (IWCMC), pp.
349–354, 2013.
[208] A. C. Bukhari and Y.-G. Kim, “Integration of a secure type-2 fuzzy ontology
with a multi-agent platform: A proposal to automate the personalized flight
ticket booking domain,” Inf. Sci. (Ny)., vol. 198, pp. 24–47, Sep. 2012.
[209] C. Porcel, A. Tejeda-Lorente, M. A. Martínez, and E. Herrera-Viedma, “A
hybrid recommender system for the selective dissemination of research
resources in a technology transfer office,” Inf. Sci. (Ny)., vol. 184, no. 1, pp. 1–
19, 2012.
[210] L. Qiu and I. Benbasat, “A study of demographic embodiments of product
recommendation agents in electronic commerce,” J. Hum. Comput. Stud., vol.
68, no. 10, pp. 669–688, 2010.
[211] N. Korfiatis and M. Poulos, “Using online consumer reviews as a source for
demographic recommendations: A case study using online travel reviews,”
Expert System with Applications., vol. 40, no. 14, pp. 5507–5515, Oct. 2013.
References
194 | P a g e
[212] B. Krulwich, “Using Large-Scale Demographic Data,” AI magazine, vol. 18,
no. 2, pp. 37– 46, 1997.
[213] S. Langer, A. Nürnberger, M. Genzmehr, T. Impact, J. Beel, S. Langer, A.
Nürnberger, and M. Genzmehr, “The Impact of Demographics (Age and
Gender) and Other User-Characteristics on Evaluating Recommender
Systems,” no. , pp. 400–404, 2013.
[214] S. Loh, F. Lorenzi, R. Granada, D. Lichtnow, L. K. Wives, and J. P. M. de
Oliveira, “Identifying Similar Users by their Scientific Publications to Reduce
Cold Start in Recommender Systems.,” in WEBIST, vol. 9, pp. 593–600, 2009.
[215] R. R. Yager and N. Rochelle, “Constraint Satisfaction Using Soft Quantifiers,”
Intelligent Systems in Accounting, Finance and Management, vol. 12, no. 3,
pp.177-186, 2004.
[216] J. A. Recio-garc and D. Bridge, “CBR for CBR : A Case-Based Template
Recommender System for Building Case-Based Systems,” In European
Conference on Case-Based Reasoning, pp. 459-473, 2008.
[217] D. Lee, “Case-based Recommendation Case-based Reasoning,” pp. 1–17,
2011.
[218] F. Lorenzi and F. Ricci, “Case-Based Recommender Systems : a Unifying
View,” In Intelligent Techniques for Web Personalization, pp. 89-113, 2005.
[219] J. Bobadilla, F. Ortega, a Hernando, and a Gutiérrez, “Recommender systems
survey,” Knowledge-Based Syst., vol. 46, pp. 109–132, 2013.
[220] F. Chaumartin, L. Talana, and U. Paris, “UPAR7 : A knowledge-based system
for headline sentiment tagging,” Proceedings of the 4th International
Workshop on Semantic Evaluations (SemEval), pages 422–425, Prague, June
2007.
[221] T. C.-K. Huang, “Recommendations of closed consensus temporal
patterns by group decision making,” Knowledge-Based Syst., vol. 54, pp. 318–
328, Dec. 2013.
[222] S. S. Anand, D. A. Bell, and J. G. Hughes, “KNOWLEDGE ENGINEERING
EDM : A general framework for Data Mining based on Evidence Theory,”
Data & Knowledge Engineering, vol.18, pp. 89-223, 1996.
References
195 | P a g e
[223] G. Adomavicius and F. Ricci, “RecSys‟ 09 workshop 3: workshop on
contextaware recommender systems (CARS-2009).,” RecSys, p. 60558, 2009.
[224] G. Adomavicius and A. Tuzhilin, “Context-aware recommender systems,”
Recomm. Syst. Handbook, Second Ed., pp. 67–80, 2015.
[225] K. Verbert, N. Manouselis, X. Ochoa, M. Wolpers, H. Drachsler, I. Bosnic, S.
Member, and E. Duval, “Context-Aware Recommender Systems for Learning :
A Survey and Future Challenges,” vol. 5, no. 4, pp. 318–335, 2012.
[226] A. Q. Macedo, C. Grande, and C. Grande, “Context-Aware Event
Recommendation in Event-based Social Networks,” In Proceedings of the 9th
ACM Conference on Recommender Systems, pp. 123–130, 2015.
[227] J. Bao, Y. Zheng, and M. F. Mokbel, “Location-based and preference-aware
recommendation using sparse geo-social networking data,” in Proceedings of
the 20th international conference on advances in geographic information
systems, pp. 199–208, 2012.
[228] P. Heshi, G. Kulkarni, K. Bidkar, and A. Oswal, “Location Aware
Recommendation System,” Imp. J. Interdiscip. Res., vol. 2, no. 6, 2016.
[229] P. Bedi, H. Kaur, and S. Marwaha, “Trust Based Recommender System for
Semantic Web, In IJCAI, vol. 7, pp. 2677–2682.
[230] C. Chen, H. Yin, J. Yao, and B. Cui, “Terec: A temporal recommender system
over tweet stream,” Proc. VLDB Endow., vol. 6, no. 12, pp. 1254–1257, 2013.
[231] D. A. DellaSala, C. T. Hanson, W. L. Baker, R. L. Hutto, R. W. Halsey, D. C.
Odion, L. E. Berry, R. W. Abrams, P. Heneberg, and H. Sitters, “Chapter 13 -
Flight of the Phoenix: Coexisting with Mixed-Severity Fires BT - The
Ecological Importance of Mixed-Severity Fires,” Elsevier, pp. 372–396, 2015.
[232] X. Zhou, Y. Xu, Y. Li, A. Josang, and C. Cox, “The state-of-the-art in
personalized recommender systems for social networking,” Artif. Intell. Rev.,
vol. 37, no. C, pp. 119–132, 2012.
[233] L. Lü, M. Medo, C. H. Yeung, Y.-C. Zhang, Z.-K. Zhang, and T. Zhou,
“Recommender systems,” Phys. Rep., vol. 519, no. 1, pp. 1–49, 2012.
[234] S. N. Revolution, “The Present and Future of Recommender Systems,”
Summer Schools Lecture, Spain, 2006.
References
196 | P a g e
[235] R. R. Yager, “Quantifier Guided Aggregation Using OWA Operators,” IJIS,
Wiley, vol. 11, pp. 49–73, 1996.
[236] J. M. Merigo and M. Casanovas, "Generalized aggregation operators and fuzzy
numbers in a unified Model between the weighted average and the OWA
operator," International Journal of Intelligent Systems, vol. 24, pp. 934-54,
2008.
[237] A. Zenebe and A. F. Norcio, “Representation , similarity measures and
aggregation methods using fuzzy sets for content-based recommender
systems,” Fuzzy Sets and Systems, vol. 160, pp. 76–94, 2009.
[238] L. M. De Campos, J. M. Fernández-luna, and J. F. Huete, "A collaborative
recommender system based on probabilistic inference from fuzzy
observations." Fuzzy sets and systems, vol.159, no. 12, pp. 1554-1576, (2008).
[239] B. Mobasher, R. Cooley, and J. Srivastava, “Automatic personalization based
on web usage mining,” Commun. ACM, vol. 43, no. 8, pp. 142–151, 2000.
[240] J. C. de Borda, “Mémoire sur les élections au scrutin,” 1781.
[241] C. Dwork, R. Kumar, M. Naor, and D. Sivakumar, “Rank aggregation methods
for the web,” in Proceedings of the 10th international conference on World
Wide Web, pp. 613–622, 2001.
[242] M. M. S. Beg, “User feedback based enhancement in web search quality,”
Information Sciences, vol. 170, no. 2, pp.153-172, 2005.
[243] R. Ali, “Pro-Mining: Product recommendation using web-based opinion
mining,” IJCET, vol. 4, no. 6, pp. 299–313, 2013.
[244] “QS World University Ranking.” [Online]. Available:
http://www.topuniversities.com/university-rankings/university-
subjectrankings/2015/computer-science-
informationsystems#sorting=rank+region=+country=96+faculty=+stars=false+
search=.
[245] L. Getoor, C. Park, and C. P. Diehl, “Link Mining : A Survey,” ACM
SIGKDD Explorations Newsletter, vol. 7, no. 2, pp. 3–12, 2005.
[246] T. E. Senator and N. F. Drive, “Link Mining Applications : Progress and
Challenges,” ACM SIGKDD Explorations Newsletter, vol. 7, no. 2, pp. 76–83,
2005.
References
197 | P a g e
[247] R. R. Yager, “On ordered weighted averaging aggregation operators in
multicriteriadecisionmaking,” IEEE Trans. Syst. Man. Cybern., vol. 18, no. 1,
pp. 183–190, 1988.
[248] G. Beliakov, A. Pradera, and T. Calvo, “Aggregation functions: A guide for
practitioners,” Heidelberg, Springer, vol. 221, 2007.
[249] M. M. S. Beg, “User feedback based enhancement in web search quality,” Inf.
Sci. (Ny)., vol. 170, no. 2, pp. 153–172, 2005.
[250] M. M. S. Beg, “A subjective measure of web search quality,” Inf. Sci. (Ny).,
vol. 169, no. 3, pp. 365–381, 2005.
[251] J. Malczewski, “Ordered weighted averaging with fuzzy quantifiers : GIS
based multicriteria evaluation for land-use suitability analysis,” International
Journal of Applied Earth Observation and Geo Information, vol. 8, pp. 270–
277, 2006.
[252] C. K. Makropoulos and D. Butler, “Spatial ordered weighted averaging :
incorporating spatially variable attitude towards risk in spatial,” Environmental
Modelling & Software, vol. 21, pp. 69–84, 2006.
[253] B. M. Rasmussen, B. Melgaard, and B. Kristensen, “GIS for decision
supportdesignation of potential wetlands,” in 3rd International Conference on
Geospatial Information in Agriculture and Forestry, 2001.
[254] M. Lee, J. Chang, and J. Chen, “Fuzzy Preference Relations in Group Decision
Making Problems Based on Ordered Weighted Averaging Operators,”
International Journal of Artificial Intelligence and Applications for Smart
Devices, vol. 2, no. 1, pp. 11–22, 2014.
[255] N. Kaji and K. Masaru, “Automatic Construction of Polarity-tagged
Corpus from HTML\nDocuments,” COLING-ACL ‟06 Proc. COLING/ACL
Main Conf. poster Sess., no.July, pp. 452–459, 2006.
[256] A.-M. Popescu, B. Nguyen, and O. Etzioni, “OPINE: Extracting product
features and opinions from reviews,” in Proceedings of HLT/EMNLP on
interactive demonstrations, pp. 32–33, 2005.
[257] L. Zhuang, F. Jing, and X. Zhu, “Movie Review Mining and Summarization
∗,” In Proceedings of the 15th ACM international conference on Information
and knowledge management, pp. 43-50, 2006.
References
198 | P a g e
[258] J. Beel, S. Langer, and M. Genzmehr, “Sponsored vs. organic (research paper)
recommendations and the impact of labeling,” in International Conference on
Theory and Practice of Digital Libraries, pp. 391–395, 2013.
[259] M. M. S. Beg, “Novel Fuzzy Queries for Searching the World Wide Web,” In
Proc. Int. Conf. on High Performance Computing (HiPC 2002), Workshop on
Soft Computing (WOSCO 2002), Bangalore, 2002.
[260] M. Ge, C. Delgado-Battenfeld, and D. Jannach, “Beyond accuracy: evaluating
recommender systems by coverage and serendipity,” in Proceedings of the
fourth ACM conference on Recommender systems, pp. 257–260, 2010.
[261] A. Gunawardana, “A Survey of Accuracy Evaluation Metrics of
Recommendation Tasks,” Journal of Machine Learning Research, vol. 10, pp.
2935–2962, 2009.
[262] W. Hersh, A. Turpin, S. Price, B. Chan, D. Kramer, L. Sacherek, and D.
Olson, “Do batch and user evaluations give the same results?,” in Proceedings
of the 23rd annual international ACM SIGIR conference on Research and
development in information retrieval, pp. 17–24, 2000.
[263] D. Jannach, L. Lerche, F. Gedikli, and G. Bonnin, “What recommenders
recommend–an analysis of accuracy, popularity, and sales diversity effects,” in
International Conference on User Modeling, Adaptation, and Personalization,
pp. 25–37, 2013.
[264] A. H. Turpin and W. Hersh, “Why batch and user evaluations do not give the
same results,” in Proceedings of the 24th annual international ACM SIGIR
conference on Research and development in information retrieval, pp. 225–
231, 2001.
[265] S. M. McNee, I. Albert, D. Cosley, P. Gopalkrishnan, S. K. Lam, A.
M. Rashid, J. A. Konstan, and J. Riedl, “On the recommending of citations for
research papers,” in Proceedings of the 2002 ACM conference on Computer
supported cooperative work, pp. 116–125, 2002.
[266] F. Ricci, L. Rokach, B. Shapira, P. B. Kantor, and F. Ricci, Recommender
systems handbook, Springer, pp. 1-35, 2011.
[267] G. Schröder, M. Thiele, and W. Lehner, “Setting goals and choosing metrics
for recommender system evaluations,” in UCERSTI2 Workshop at the 5th
References
199 | P a g e
ACM Conference on Recommender Systems, Chicago, USA, vol. 23, p. 53,
2011.
[268] F. Hernández delOlmo and E. Gaudioso, “Evaluation of recommender
systems: A new approach,” Expert System with Applications., vol. 35, no. 3,
pp. 790–804, 2008.
[269] P. Cremonesi, R. Turrin, E. Lentini, and M. Matteucci, “An evaluation
methodology for collaborative recommender systems,” in International
Conference on Automated solutions for Cross Media Content and Multi-
channel Distribution, AXMEDIS, pp. 224–231, 2008.
[270] J. Vinagre, A. M. Jorge, J. Gama, "Evaluation of recommender systems in
streaming environments," arXiv preprint arXiv:1504.08175. Apr 30, 2015.
[271] J. Thom and F. Scholer, “A comparison of evaluation measures given how
users perform on search tasks,” in Proceedings of the 12th Australasian
Document Computing Symposium, Melbourne, Australia, December 10,
(ADCS), pp. 100-103, 2007.
[272] A. C. Nielsen, “Nielsen: Global consumers‟ trust in „earned‟ advertising grows
in importance,” Bus. Wire, 2012.
[273] R. Ali and M. S. Beg, "Modified rough set based aggregation for effective
evaluation of web search systems," In Fuzzy Information Processing Society,
NAFIPS, Annual Meeting of the North American, pp. 1-6, 2009.
[274] M. M. S. Beg and N. Ahmad, “Subjective Enhancement and Measurement of
Web Search Quality,” In Enhancing the Power of the Internet, pp. 95-129,
2004.
93 | P a g e
Chapter 5
Feature based Opinion Mining Approaches for
Book Recommendation
5.1 Introduction:
The previous Chapter deals with the recommendation issues of the top books to
universities‟ students using suggestions made by their institution, and keeping the
university prescribed syllabus as a foundation for the problems. The dataset was the
recommended books in top Indian universities. The main idea was to incorporate link
mining techniques with the help of robust aggregation methods for sorting the top
books amongst all. In this chapter, the dataset of books was kept as it is, but we have
modified the approach in the recommendation process of top books. Instead of
directly applying weighted or un-weighted aggregation method to the books
prescribed in the syllabus of universities, we have collected the customer reviews
available at different online retailer sites of books. Thus, opinion mining which is a
type of web content mining is exploited in addition to link mining which has been
used in previous chapter. Collection of these reviews is the first step in opinion
mining. These reviews are found in different form; usually reviews are open views of
a user in any language, unformatted and unstructured [255]. Hence, deciding about a
product from the available online user‟s unstructured opinions is a tough task, though
very interesting [256].
Once the opinion or reviews of the users are obtained, next step is to enforce some
methodology for converting these reviews in some operational form so that we may
process the reviews to get acquaintances with the features of the item for which all
these opinions are made. The feature extraction is a very important aspect of opinion
mining process. If the appropriate features are extracted from varieties of the options
for the targeted item, i.e. items whose top products are to be recommended, the
reviews can be assessed in the perspective of these features, which, then, may lead us
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
94 | P a g e
to decide about the products and what are the feedbacks of the user according to their
experiences with it.
The sentiments of the users are attached with reviews. By reviews, one can know
the experiences, emotions and sentiments hidden in their words. The reviews may
have some positive words about a product or negative. The main task is to find the
sense of users from the reviews about the extracted features of the items, whether they
are positive or negative [255], [257]. Generally, for analyzing about the positivity or
negativity of a review, some pre-determined set of words are examined. Words like, “
well, fantastic, written by an experienced team, exactly what I needed, good job,
covering everything you need to be aware of, especially appreciated, awesome book,
classic, well written, highly recommend ” etc. are treated as positive terms. Examples
of negative words include, “ real disappointment, Not a general discussion, fluffy,
worst business, copied, biggest failing, worst, time consuming, bad, not
recommended,” etc. are termed as negative comments. It is important to observe the
reviews with human intelligence for interpreting the true sense of the reviewers as the
reviews are representations of human emotions. Sometimes it does happen that things
appear to be different then what they are. Let us consider the following examples:
Example 2.1:
Example 2.2:
In the above example 2.1 and example 2.2, there are positive words which can be
realized as a positive sentiment of user in favor of the books. The terms like „do your
students a favor‟ in first example, and „highly recommend‟, „better option‟ in second
example seem to be representing a positive sense; however it is not the case, as all
above positive terms are used in a negative sense. Thus, keeping only the positive and
negative aspects of the terms and processing on this much of knowledge may lead to a
wrong conclusion. This is the point which is not addressed by several opinion mining
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
95 | P a g e
algorithms. We have proposed algorithms that take into account the reciprocal-
meaning interpreting terms and employ human intelligence for a final decision.
5.2 Customer Reviews
As the customers reviews serve as a base for the recommendation of books in the
feature based opinion mining approach, we have collected reviews from highly rated
online merchandiser of books worldwide and different sites that allow the users to
present online reviews. The list of the sites from where the reviews are obtained is
listed below.
1. http://diestel-graph-theory.com/reviews.html
2. http://www.amazon.in/Concrete-Mathematics-Foundation-Computer-
Science/dp/0201558025
3. http://www-cs-faculty.stanford.edu/~uno/gkp.html
4. http://www.flipkart.com/concrete-mathematics-foundation-computer-science-
2nd-english/p/itmdx9se8fvvfph8
5. http://www.goodreads.com/book/show/1041923.Discrete_Mathematics_for_Co
mputer_Scientists_and_Mathematicians
6. http://www-groups.dcs.st-and.ac.uk/history/Extras/Harary_books.html
7. https://dzone.com/articles/compilers-principles
8. http://www.cambridge.org/ae/academic/subjects/computer-
science/programming-languages-and-applied-logic/logic-computer-science-
modelling-and-reasoning-about-systems-2nd-edition
9. http://shop.oreilly.com/product/9781565924536.do#PowerReview
In various cases authors have their own web pages that link to the resources of
reviews from the users. Though there are good numbers of books for which online
shopping portals have users review, however, there are some reviews that are
published by the publisher and not by a common user. Since, we are interested in user
reviews, the reviews from the merchandisers, magazine editors, writers and any
sources which may have biasness in reviews, have not been considered.
5.2.1 Issues while handling Online Reviews:
Several problems are encountered while handling online reviews. One of the problems
is of languages other than English in which reviews are available. Few examples of
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
96 | P a g e
reviews other than English language found at several sites are shown in Figure 5.1,
5.2, 5.3 and 5.4 respectively. While handling these reviews, the Google translator is
used for translating into English so that the reviews can be processed through
proposed algorithms.
Figure 5.1: Demonstration of a review in Spanish
Figure 5.2: Demonstration of a review in Russian
Figure 5.3: Demonstration of a review in Portuguese
Figure 5.4: Demonstration of a review in Greek
However, there are various books for which not enough number of reviews is
found, even there are the books which have not a single customer review. An example
is shown in Figure 5.5.
Figure 5.5: Screenshot displaying no reviews
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
97 | P a g e
Figure 5.6: Architecture of Book Recommendation using meta searching
5.3 Feature Extraction and Selection
In[1], authors have used the technique to categorize the features of the books for its
recommendation using meta searching as shown in Figure 5.6. The authors have
queried with key words on different search engines for finding the top books on the
discipline concerned. Key words like “books on the „specific course‟ ” for more than
10 disciplines of the computer science are passed. Forevery coursethe query seems
manifested systematically, e.g., let us consider books on compiler design; the query is,
"books on compiler design". The names of the books that appear in top 100 links are
stored with the help of Search Engine Optimization (SEO) tools.
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
98 | P a g e
Again, the authors try to extract the features from opinion. When the opinion
extraction is complete, ranking of the items (books in this case) based on scores are
performed. The feedbacks from the customers can be positive, negative or comments
may be neutral that contains neither negative nor positive terms in reviews. We have
analyzed the reviews and categorized several features for providing a better
understanding to the customers of books. The features are extracted from reviews on
the basis of the adjectives and indicative terms used by the users. Let us take an
example.
“The book has plenty of material to explore, written in a good manner but the
cost is too much to buy.”
The text expressed by the readers tells about the book features. We can extract
from the review that users are talking about „study material of the books‟,
„understandability of the contents‟ and „price‟. Thus, three features are extracted. In
such manner we have found only seven important features which are considered in the
recommendation procedure. These features are presented to users [37]. The users
were asked to give their feedback on importance of features. The users‟ feedbacks
were stored and analyzed to get an idea for creating bases to perform feature based
opinion mining.
The precision for the extracted feature is calculated and the observed precision is
considerably high which indicates that users have agreed on these features
unconditionally in a major proportion. The precision is the proportion of the users
favoring the features to total number of users involved in the procedure.
The categorized seven features play a vital role and serve as a foundation for the
process of feature based opinion extraction in the book recommendation approach. The
features and their explanations are discussed below.
Frequency of Occurrence in SERP: Occurrence means the multiple appearances of
the same books on single query. When users would like to see various options of the
books for a particular topic, they usually browse to search engine sites. There are
various books that have multi occurrences in the form of different link in any search
engines. Also, there may be books which occur in more than one search engines in
very first appearance of the Search Engine Results Page (SERP). This characteristic is
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
99 | P a g e
termed as „frequency of occurrencesin SERP”. Figure 5.7 shows the example of a
SERP.
Figure 5.7: Screenshot of Seacrh Engine Result Page for books on Artificial Intelligence
Useful Content: sometimes the books are well written, however the content is
either not sufficient or the authors would have merged all content together,
which make the reader irritating. In contrary, if the contents are useful, though
short and precise, the readers get attracted towards the book and eventually
show a love towards it.
An example of user review from amazon.com that favors our argument is
shown in the Figure 5.8.
Figure 5.8: Customer review expressing the views about content
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
100 | P a g e
Rating: rating is done by the user through various websites. The collective
rating of users on a particular books advocates about the user preferences
towards a book. Few of the listed sources in the previous section consist of
rating along with reviews. The higher rating of the user indicates the popularity
of a book amongst readers. Most of the techniques that support the philosophy
of recommendation approaches use only rating based recommendation.
However, we have incorporated rating as well as 6 other characteristics in our
recommendation process.
Understandablility: The content of the book, though how reach, should be
written in an easy-to-understand way, i.e. it should be understandable. The
users in their reviews use to write that book has a lot to give but not able to
understand the way writer has presented. The example words for this feature as
well as other features are listed in Table 5.1andFigure 5.9shows its importance
interpreted from user review.
Figure 5.9: Review example of „understandability‟ feature.
Physical Attributes: the physical attribute is concerned with the quality of the
pages, hard cover, and print quality etc. of the books. An example from screen
shot of a user review available at amazon.com is shown in Figure 5.10.
Sometimes it is found that the in spite of books being written brilliantly has low
value in the eye of the customer due to its various physical attributes. The
following example is an illustration taken from Amazon.
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
101 | P a g e
Figure 5.10: Review representing importance of physical attributes
Market Availability: the services of the book associated with publishing
houses, its market strategies and other related features which are reflected by
user reviews are placed in this feature. Also, it includes availability of the
reviews. As it is discussed in the previous section, there may be some books
which do not have even one review (Figure 5.5); hence their market
availability will have fewer values.
Price: This is the obvious feature of the book, a student is interested in. An
example of real user review emphasizing the importance of the considered
feature is shown in Figure 5.11
Figure 5.11: Review representing importance of Price
Table 5.1: Features and related review terms
Symbols Features Adjectives indicating features
SUC Useful Content
Content, material, helpful, useful,
advantageous, cover, Time waste,
creeping all together
SU Understandability
Hard to understand, easy to understand,
convey, fantastic written, Explained very
well,
SA Market Availability Available, sold,
Quick delivery
SR Rating Star, full rating, rating, rated,
SP Price Cost, price, worth, cheaper, costly, not
worthy for the amount
SPA Physical Attributes
Cover page, page quality, The cover of
the newer edition (2006) is pretty dull,
etc.
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
102 | P a g e
The above seven features can be expressed by number of ways and hundreds of words
in English. The specific terms which are used to express the sentiments have been
highlighted an above examples. The common terms which are usually manifested by
users to convey the characteristic of the books are indicated in Table 5.1.
5.4 Scoring Technique for Extracted Feature
The recommendation approach is basically based upon two step scoring process. First
step is to find the score for all seven (7) features discussed in section 5.3 associated
with each book and second, assigning weights to the feature. In section 5.4.1, the
calculation of opinion score is described which is aided by several algorithms.
5.4.1 Opinion Score Calculation
The opinions are combination of positive and negative sentiments, expressed by words
conveying similar meaning. There are various suggestions made by the researchers for
a detailed list which clearly identify whether the word involved in the sentence has a
positive sense or negative.
However, we have discussed earlier that only counting and calculating positive
words and negative words may mislead in finding exactly what the sentences meant
for? That is why we have designed a calculation method which involves the positive
and negative words along with those words which seems to be positive but in
association with some other reciprocal phrases it covey the almost opposite meaning.
These terms are said, “Reciprocal terms”. S_pw and S_nw represent scores for positive
and negative words respectively.
Algorithm 5.1: Score calculation of Positive Words
_ ) – (( )
_=
_ * 1.5_
i
N pw N pr N hpS pw
r
N_pw = No. of Positive words
N_pr = no. of positive words with reciprocal meaning
N_hp = no. of highly expressible positive words
_i
i
Nr
N R
_ No. of reviews considering feature 'i'iN R
N = Total no. of reviews extracted
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
103 | P a g e
Algorithm 5.2:Score calculation of Negative Words
_ ) – ( _( ) =
_ * 1.5_
i
N nw N nr N hnS nw
r
N_nw = No. of negative words
N_nr = no. of negative words with reciprocal meaning
N_hn = no. of highly expressible negative words
The explanations and examples for the positive words, negative words, reciprocal
terms with positive and negative words, highly expressible positive and negative
words are discussed below.Examples of positive, negative and reciprocal words are
explained in the next sections.
5.4.1.1 Positive words:
The sentiments which are straight forward and conveying positive meanings are
categorized and placed in positive words. The number of positive words is indicated
by N_pw. An example for positive review is underlined.
Example 5.1a “Positive review”: “I taught a couple of classes from the first edition
of this textbook, and my students did fairly well. On the whole, they were able to
understand the material and solve the homework problems. I certainly wouldn't mind
teaching a class on this subject from the second edition as well, which I feel is a mild
improvement over the first one.”
Example 5.1b “Positive review”: “The Chapteron finite automata is excellent. And
the material on context-free languages is thorough and well written. So is the
introduction to Turing machines. Of course, the book then spends a fair amount of
time on recursive function theory. That is exactly what I want it to do. And I think the
Chapteronunsolvability, starting with the Halting Problem, is excellent.”
5.4.1.2 Negative words:
The sentiments which ditrectly convey negative opinion from the customers are
categorized and placed in negative words. The number of negative words is indicated
by N_nw. few examples for negative review are illustrated below.
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
104 | P a g e
Example 5.2a “Negative review”: “Apparently, the only way to understand this book
is by having gotten your PhD in the 1950's. Completely incomprehensible, stilted, and
pompous, this book is the long sought after cure for insomnia. If you are a professor,
please do not choose this book for your class. If you are a student, pray”.
Example 5.2b “Negative review”: “So, you're basically paying anywhere from $100
to150 for the newest cover art and 25 pages. Don't waste the money.”
Example 5.2c “Negative review”: “This book is horrible. Please read some other
book which explains data structures in plain English. This book has two sections
called array data types and array data structure and both of them have pretty much
the same stuff written in a different, complicated way. Each is four pages long as well.
Not a good read and I will definitely not recommend it. If you want to memorize
definitions this might be the book but if you want to understand concepts, read
something else.”
We have proposed an algorithm for scoring to consider the value of positive and
negative words accordingly. The method is shown in Algorithm 5.3.
The above procedure helps in finding the negative orientation contained in a review.
The method provides in analysing the text and depending upon the orientation of the
sentiments expressed by users, products are scored.
Algorithm 5.3: Search the positive words in the document (opinion).
N_pw Count (positive words)
Search the negative words in the opinion.
N_nw Count (negative words)
If (N_pw>N_nw)
{
OO positive
}
else
OO negative
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
105 | P a g e
5.4.1.3 Reciprocal terms
In the above examples from 5.1a – 5.2c, the sentences are clearly stating positive or
negative aspects of the features associated with the books. However, there are the
situations when a word which may be grouped as positive words but would be used in
conveying its reciprocal sense, i.e. negative and vice-versa. These terms, as discussed
in section 5.4.1, are treated separately in calculating feature scores and called as
„reciprocal terms‟. Let us consider an example.
“The derivations, definitions and everything are written in a very easy and
perceivable way. But again the same warning, don't buyif you are a beginner”.
In the above example the terms like very easy, perceivable etc. represent positive
words however user is clearly forbidding from the purchase of the books. Few other
examples of reciprocal terms include:
Example 5.3a “Reciprocal terms”: “I would say it is a fine undergrad book, but
probably not the best choice for grad level studies.”
Example 5.3b “Reciprocal terms”: “This is a nice read for someone entering the
software architecture domain. However, I would not recommend it as a reference
book for a software architect.”
Example 5.3c “Reciprocal terms”: “Deserve one of the bestpoorly written for CS
theory.”
Example 5.3d “Reciprocal terms”: “Nearly all of the writer's explanations are
lacking at best. There are practically no examples to help you understand what the
writer is trying to convey. The answers in the back of the book seem to only be for the
easiest questions. Many of the proofs are incomplete as the writer intends for you to
come up with them in exercises without adequate explanation.
In short, if you have to use this book, I`m sorry”
Example 5.3e “Reciprocal terms”: “Not a book I would recommend.”
In the above examples we would easily recognize there are the supportive words in
the review which may seem in the favor of book, however they are not at all. Like in
Example 5.3e, „I would recommend‟ is an obvious positive and strong word in the
favor of book; however it is used totally in a negative sense with „not‟ in start of the
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
106 | P a g e
sentence. For understanding reciprocal terms one should have a good ground of
English grammar. There are the terms which confuse the meaning of a sentence. It
starts with a negative word and tries to convey its reciprocal meaning, and vice-versa.
Let us consider an example with „not only but‟, as it may convey dual meaning. “He‟s
not only funny, but also he‟s intelligent.” The „not‟ is usually classified as a negative
word, however „not only‟ is used to exhibit a reverse feeling.
Example 5.3f “Reciprocal terms”: “The book is very comprehensive and explains the
needed background to allow readers to not only use metrics well, but to understand
the limitations of metric”.
Sometimes readers compare a book with other and use terms to discuss the
differences, however if a true measure is not applied the opinion interpreter would
convey the meaning incorrectly.
Example 5.3g “Reciprocal terms”: “What a terrible book. Though it's the
cornerstone of many CS undergrad algorithm courses, this book fails in every way. In
almost every way, Dasgupta and Papadimitriou's "Algorithms" is a much better
choice.
Algorithm 5.4: Finding Opinion Orientation “OO” using reciprocal terms
Search for the reciprocal terms (rt) in the sentence.
If (rt ϵ Sentences)
{ Find the OO of the sentence before rt;
If (OO before „rt‟ is positive)
{
OO negative
N_pr Count (rt)
}
else
{
OO positive
N_nr Count (rt)
}}
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
107 | P a g e
Algorithm 5.5: Finding Opinion Orientation for „not only‟ phrase
Is (((not only) ϵ (Sentence))∧ ((but also) ϵ (adjac. (not only) ∨ Near5 (not only))))
If (pw ϵ ((adjac. (but also)) ∨Near2 (but also)))
OO positive
Else if
{
(nwϵ ((adjac. (but also)) ∨Near2 (but also)))
OO negative
}
The term „better choice‟ in the above example is not used for the book but for some
another one, and observing these terms is very important while dealing with usrs
sentiments through their reviews. We have proposed reciprocal terms finding
algorithm to enrich the process with a correct measures of understanding sentiments.
However, there are few words which are often used as reciprocal terms. These are
although, however, nevertheless, on the other hand, still, though, yet, but, etc.
As in Example 5.3a and 3.3b the meaning of sentences before reciprocal term is
changed by the sentence used after reciprocal terms.Keeping in view the issues
mentioned in the above examples, we have proposed the following algorithms to
tackle with. Opinion Orientation „OO‟ of the reciprocal terms is found from the
proposed procedure which will boost the reviews interpretation. Algorithm 5.5
presents the considerable inclusion of reciprocal meanings in a text. To deal with „not
only‟ phrase, the algorithm is suggested below.
5.4.1.4 Highly expressible words
By highly expressible word we mean those words which do have extra exclamations
and have been assigned a high score than simple affirmative word. E.g. “One of the of
the finest computer science textbooks I've ever read, and I've read hundreds.”
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
108 | P a g e
Algorithm 5.6: Counting Highly Expressible positive and negative words
If (highly expressible words)
{
If (OO positive)
{
N_hp Count (positive words)
}
Else
{
OO negative
N_hn Count (negative words)
}
}
With the help of six algorithm presented in the chapter, the sentiments are associated
with scores. Each user‟s sentiments expressed through their reviews are numerically
assigned values, which we call as opinion score. After finding the respective scores
for positive and negative words, the conclusive score „S‟ is calculated by;
( _ ) _ ) ---- (5.1 )(S S pw S nw
5.4.2 Weight Assignment to Features
In the section 5.3, it is discussed about how the features are selected which are further
served a base for finding opinion orientation towards these features and scoring the
books accordingly. We have already discussed in the section 5.1 that the important job
to be performed is score calculation and weight assignment to features. Section 5.4.1
describes how the scores are calculated for associated features of books. Here, we give
the details of the weights distribution scheme to the extracted features.
We have considered WO,WUC,WR, WU,WA, WP, and WPA as weights assigned to,
Frequency of Occurrence in SERP, Useful Content,Rating, Understandability,
Market Availability, Price and Physical Attributes, respectively. In the previous work
[1], the authors have used different ranges of the weights depending upon the use of
the features. However, they have selected the features without any scheme and they do
not have any defined method for assigning weights.
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
109 | P a g e
Algorithm 5.7: Weight assignment to features
n number of users participated in the feedback
r number of users with „true positive‟ feedback for feature „i‟
Tf total number of features extracted; Tf = 7
fi ith
extracted feature
Sfi (r/n)*100
Find (maximum amongst Sfi , for 1≤ i ≤ Tf )
Smax Max (Sfi)
WfiSfi / Smax
}
Here, the weights are assigned according to the importance that the users have
suggested in their feedback, as discussed in section 5.3. An algorithm for weight
assignment is depicted in Algorithm 5.7. The final score for a book is calculated as
follows;
Let si is the score of ith
feature associated with a book, and wi is the weight assigned to
feature „i‟. The final score (F.S) of the books is calculated as;
7
i
1
----- (5.2)i
i
FS w s
The books are sorted according to FS value which gives a ranked list of books. The top
books are then recommended for users.
5.5 Results and Discussions
Here, in this Chapterthe direct opinion of users are taken from online review sites
and these reviews are processed to ascertain the preference of the users. According to
the users‟ manifest, the books are recommended. Since, we have a total of 10 different
courses consisting of 158 distinct books, as discussed in Chapter 3. A total of 0.4587
× 104 reviews are obtained for these books from the above sources. There are seven
features regarding books which have been extracted from reviews, and total numbers
of 100 users are considered for feedback collections.
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
110 | P a g e
Table 5.2: Precision of extracted features
Extracted
features
Frequency of
Occurrence
in SERP
Useful
Content Rating Understandability
Physical
Attributes
Market
Availability Price
Precision 0.84 0.92 0.78 0.79 0.67 0.73 0.71
Figure 5.12: Precision of Extracted Features
The precision of the extracted features is also calculated. The precision is the
proportion of the users favoring the features to total number of users involved in the
procedure. The precision value of their feedback is tabulated in Table 5.2, and
pictorially represented in Figure 5.12.
It is evident from above table that all the features have higher value of precision.
The worst precision is recorded for feature „physical attributes‟ which is 0.67.
The weights are assigned to these seven features as described in algorithm 5.7. The
extracted features are presented before users to give their feedback on the concerned
features, whether they consider it important or not? The user feedback helps us in
deciding on what features the books recommendations would be made. To understand
how weights are assigned to features, let us take precision of useful content. From
Table 5.2, we have p (useful content) = 0.92. Weight assigned to useful content „Wuc‟
is given by;
( ax)
( )
m
ucp usefulcontent
pW
Wuc = 1.
It is evident from the table, „useful content‟ is the most valuable feature obtaining
1, maximum weight, i.e. if for a book the opinion score of feature useful content „Suc„
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
111 | P a g e
is high it will have a greater impact in its recommendation. Similarly, WU = 0.858696
which employ that opinion score of „understandability‟ will have 85% weightage in
scoring of users sentiments.
The minimum weight is obtained by „physical attributes‟ which has lowest
precision „.67‟. Thus, if a book is best reviewed from its physical attribute point of
view but is not attracting the users for other highly weighted features like „useful‟
content, the book well have a less score, and hence would get a lower ranking. The
weights obtained for respective features are given in Table 5.3.
A special feature, “Occurrence in SERP” is also included which gives the
importance of the books in search engine. The frequent and top ordered appearance of
a book on SERP clearly supports its high value in the eye of users, as most hit are
made on it. So, overall only the books which have high values from all the aspects
will be sorted and ranked. These ranked books are recommended to users.
Table 5.3: weights distribution of features
Features Weights
Frequency of Occurrence in SERP 0.913
Useful Content 1
Rating 0.8478
Understandability 0.8586
Physical Attributes 0.7282
Market Availability 0.7934
Price 0.7717
Table 5.4: Score calculation example
Frequency
of
Occurrence
in SERP
Content Understandability Rating Physical
Attributes
Market
Availability Price
N_pw 4 9 12 45 0 7 5
N_pr 0 0 0 0 0 0 0
N_hp
0 0 3 0 0 0 0
N_nw 0 5 3 39 0 2 4
N_nr 0 0 0 0 0 0 0
N_hn 0 0 2 0 0 0 2
N_ri 1 11 9 84 52 9 9
N 1 84 84 84 84 84 84
ri 1 0.13 0.11 1 0.62 0.11 0.11
S_pw 4 1.18 1.77 45 0 0.75 0.54
S_nw 0 0.65 0.64 39 0 0.21 0.75
S 4 0.52 1.13 6 0 0.54 0.21
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
112 | P a g e
The number of different books for each course has been discussed in Chapter 3. The
detail is given in Table 3.14. We have 17 different books of discrete mathematics.
Total 364 reviews are obtained for all books on discrete mathematics. This implies a
book has average 20 reviews. The calculation which is shown in Table 5.5is
performed for all the reviews of each and every book. Final score (F.S) then is
calculated. The books are ranked in order of scores achieved based on score
calculation of opinions.
Table 5.5: Example of Final Score Calculation
Features Weights Score „S‟
Frequency of Occurrence in SERP 0.913043 4
Useful Content 1 0.52
Rating 0.847826 1.13
Understandability 0.858696 6
Physical Attributes 0.728261 0
Market Availability 0.793478 0.54
Price 0.771739 0.21
0.913043*4 1*0.52 0.847826*1.13 0.858696*6
0.728261*0 0.793478*0.54 0.771739*0.21
10.8729
(
3
)
FS
FS
Table 5.6: Top 10 ranked books of all the courses using Opinion Mining Technique
Rank
position Books of different courses
1 TOC.1 CD.1 AI.2 SE.2 DB.1 DS.1 CN.16 DM.1 CG.7 OS.2
2 TOC.4 CD.4 AI.11 SE.4 DB.2 DS.9 CN.2 DM.3 CG.4 OS.1
3 TOC.9 CD.2 AI.4 SE.14 DB.11 DS.4 CN.1 DM.4 CG.10 OS.3
4 TOC.10 CD.3 AI.12 SE.5 DB.3 DS.2 CN.6 DM.2 CG.11 OS.4
5 TOC.8 CD.5 AI.1 SE.10 DB.4 DS.7 CN.8 DM.8 CG.9 OS.5
6 TOC.3 CD.9 AI.15 SE.6 DB.5 DS.15 CN.5 DM.9 CG.6 OS.6
7 TOC.2 CD.6 AI.8 SE.8 DB.12 DS.10 CN.7 DM.10 CG.12 OS.7
8 TOC.6 CD.7 AI.20 SE.7 DB.8 DS.3 CN.11 DM.14 CG.14 OS.11
9 TOC.11 CD.10 AI.3 SE.3 DB.13 DS.11 CN.3 DM.13 CG.15 OS.9
10 TOC.7 CD.8 AI.5 SE.9 DB.6 DS.6 CN.4 DM.15 CG.17 OS.10
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
113 | P a g e
On the basis of the scores, ranking of books for all the courses are shown in Table
5.6. The ranking of these books based on different soft computing techniques are
given in Table 4.21 to 4.29. The detail discussions of those rankings have been made
their. In next chapter, the different soft computing techniques will be discussed and
their methodology will also be described. Here, we just highlight how the
recommended books behave with respect to different approaches used for book
recommendation. On comparing the ranking based on Opinion mining technique
(OMT) with those of soft computing, it is observed that books on theory of
computation has same recommendation from all the techniques.
Further, TOC.11 is ranked in top position using OMT; however, PAS and ORWA
techniques do not recommend it in top positions either. SE.2 is ranked first by all the
procedure adopted in Chapter 3 and it has secured first positions using opinion mining
technique, presented in this Chapter also. The usual similarity of OMT with ORWA
and PAS can easily be noticed as OS.2 which has secured first rank in the ranking of
all these techniques. OWA with quantifier at least half has also the same
recommendation. However, OWA with „as many as possible‟, and „most‟ quantifier
differ in their final recommendation for first position.
The configured discussion on various parameters for these recommended books are
discussed in details in the Chapter 6. Further, interpretation of results from various
aspects are also elaborated and exhaustively analyzed in the chapter.
5.6 Summary
In this Chapter the book recommendation methods are discussed. The proposed
method is based on opinion mining. The technique for extraction of opinions and
selection of features are discussed for various leading book retailers. The feature
based recommendation helps users in exploring the items of their preferences as it
ease the process of selecting the desired products when users are aware of their
requirements and specification of the items.
It has also been shown that the books features which are extracted have high
precision. The high value of precision shows the strength of the feature selection
procedure.
A recommendation technique influenced by experts‟ opinion has been conferred in
Chapter 3 and 4. The same data set is used here in this chapter. But instead of experts‟
Chapter 5: Feature based Opinion Mining Approaches for Book Recommendation
___________________________________________________________________________________
114 | P a g e
suggestion of books, users‟ reviews are considered. The ranking on the basis of
different approaches are discussed. The details of the data set and selection of books
is discussed extensively in the Chapter 3 and 4.
The method is based on opinion mining, as discussed above. It is obvious; the
user‟s reviews will help in recommendation as most of the users would seek to know
the opinion of other users prior to buy an online product. For exactly evaluating the
best one, the comparison of different parameters is needed and an exhaustive
approach is required. In the Chapter 6, a comprehensive approach is discussed that
provides a guideline for the evaluation of RS and has the provision to avoid the
inclusion of insincere users from evaluation process while taking feedback from them.
The sixth Chapter also compares the results of the explicit feedback recommendation
approaches proposed in Chapter 3, Chapter 4 and Chapter 5.
115 | P a g e
Chapter 6
Evaluation of Recommender Systems
6.1 Introduction:
The proliferation of the Internet has enlivened online shopping. The increased interest
in online shopping has caused emergence of a large number of online merchandisers.
Due to the significant increase in online merchandiser, there is an enormous increase
in the number and varieties of products being sold on the Web [3], [4]. Consequently,
for a buyer, finding products of his choice in online shopping has become a tough and
tedious job. Merchandisers provide recommendation for buyers to help them in
getting the products of their choice. The recommender system also helps the
customers a lot by reducing the time spent in the exploration of different products of
their choice. On the other hand, it fulfills the merchandisers‟ interest of being at top
by exposing their competence in business, as it is likely that a good recommender
system will enhance the marketing strategy of the merchandisers and help them to
attract the customers [2], [258].
The ultimate goal of a recommender system (interchangeably termed as
recommendation system) is to satisfy the user [213]. User satisfaction depends upon
fulfilment of their needs. Different users have different requirements. Thus we should
opt for a recommender system that can identify the different users‟ requirement and
can predict items to them according to their needs. However there are other important
aspects while designing a recommender system along with the consideration of user
satisfaction. The system should not exhibit errors that can irritate the users while they
purchase, otherwise a user may never come again for purchasing. The evaluation of a
recommender system helps us in judging these errors and designing an application
that may fulfil the user satisfaction. The recommender system performs the job of
providing users better option for their purchase, the evaluation of a recommender
system judges its performance and helps in employing the modification in the system
according to the user‟s need and as per the shortcoming encountered while
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
116 | P a g e
performance assessment is performed. Thus, it is necessary to have an adequate way
of assessing the recommendation approaches which may be achieved by a sound
evaluation process.
This Chapteraims at exploring the techniques to reduce the flaw in Recommender
Systems (RS) so that users can get most appropriate recommendation for their online
shopping. Recommendations based on fake review are usually biased. The biased
recommendation made by online shopping portals may have a negative impact over
the customers and it may lead to further reluctance of their purchase from the same
online shopping site. Thus, if a RS can be designed in order to reduce the factors
which can affect the users purchase negatively, eventually it will help online
merchandiser to boost their business as well as customers would be provided with the
suitable products of their choice. Since RS are meant to provide customers with ease
and satisfaction, hence user feedback can be a key component in evaluating the RS
and checking whether it is up to the mark or not? The proposed approach provides a
platform to evaluate the RS on the basis of user feedback. In addition to this, it also
examines the user‟s sincerity and put preference criteria, which in turn must overcome
the biasedness and fake review problems in designing RS.
User feedback is one of the strongest bases for evaluating a recommender system.
However, there are two types of user feedback.
Implicit user feedback
Explicit user feedback
Implicit user feedback requires the detail of user behavior which in turn gives their
trend and inclination towards purchase. Explicit user feedback needs feedback from
users described explicitly. In this chapter, we have extensively discussed evaluation
measure based on both types of feedback. A recommender system which is designed
to recommend electronic items like, laptop, tablet, smart phone, headphone and
printer [243] is evaluated using implicit user feedback. As, these items include
reviews from a common user, which may be biased or casual. We have suggested an
evaluation approach that checks the user sincerity and followed implicit feedback
strategy. On the other hand, in this work, we have also recommended books and
evaluated the recommendation technology. The difference lies in sources from where
feedbacks are taken. For books, we have explicitly taken feedback from experts, who
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
117 | P a g e
are computer professionals and graduates. The pros and cons of books for any
specified topic can never be judged by a common man but by experts. Thus experts
ranking have become basis for the evaluation of book recommendation approach.
Hence, there is no need of sincerity checking and finding the criteria of preference,
instead, we can simply ask direct ranking by the user. This may be treated as a
standard ranking. Further, it may serve as a basis for the evaluation of the system. The
detail of the explicit feedback based evaluation is elaborated in next chapter; however,
later in this Chapter a brief discussion on the topic is made.
With the brief discussion of explicit feedback and its differences with implicit
feedback which has been discussed above, it is understood that if the users are
authentic, reliable and experts of the field concerned, the explicit feedback is
preferred. In Chapter 5, we have comprehensively discussed about the evaluation of
recommender system where the feedbacks are implicitly taken from the users. The
users considered for the evaluation procedure were common people who use the
products and need not to be an expert for presenting their opinion on the concerned
items. However, there could be events or items for which general users‟ behavior for
purchasing may not lead to exact conclusion, like books, institutes, conferences, etc.
Since, in our proposed work, we have opted specifically book for recommendation,
which further can be extended to any product of various domain. As to assess the
book, one need to be an expert and the expert‟s opinion is something which can be
relied upon. Thus, the experts of the different subjects from all over the globe who is
somehow related to Indian education system are considered and approached for their
recommendation and reviews on the specified topic of the books concerned. Their
feedbacks are kept to serve as a basis for the evaluation of the proposed system which
is presented in Chapter 3, 4, and 5.
Once we get the ranking from the experts, the evaluation metric is needed for the
evaluation of the performance of the proposed work by comparing the ranking
obtained by suggested book recommendation approaches (BRA) and the ranking
given by experts. P@k, Mean Average Precision (MAP), FPR@10, FNR@k, Root
Mean Square Error (RMSE), Modified Spearman Rank Correlation Coefficient and
Mean Absolute Error (MAE) are used as evaluation metrics for the assessment of the
approach. The brief explanation of these evaluation metrics is discussed in section
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
118 | P a g e
6.3. The corresponding values of the respective parameters with adequate discussions
are presented in section 6.4.
6.2 Previous Evaluation Studies
As the approaches for recommendation of products, books, research articles, etc. are
becoming increasingly popular, the evaluation of the recommender system is
becoming very important so that it can be assessed that which approach should be
adopted for recommendation as per the need of the users.
Prior to evaluation of recommender system it is necessary to know how a
recommender system can be measured. However consensus is the basis for judging a
good recommender system and the methods adopted for evaluation of recommender
system[51].
Initially, researchers emphasized on accuracy as a measure for the evaluation of
recommender systems. They used accuracy based on MAE as an evaluation measure
[259], [260], [186]. Later, aspects other than accuracy were suggested, as it is never
mandatory that an accurate system is always good [186], [261], [262]. The author in
[259] focused on the quality of the recommender system instead of just finding the
accuracy of mathematical approach or algorithm. The quality of a good recommender
system cannot be judged by finding predictive accuracy only but it should meet the
user satisfaction which is the ultimate goal of a recommender system. In literature,
several features that constitute a good recommender system are discussed. Authors in
[9] discuss user satisfaction and satisfaction of recommendation provider along with
accuracy as main features for an adequately reliable recommender system. In [42],
[51], [259] the authors have suggested features like coverage and serendipity are also
important factors. In [186], [263] response time of recommendation is considered as
an important factor. However, presentation is also important factor that can enhance
the user satisfaction [264], [265].
What are goals of RS and how to choose metrics for evaluation is studied [266] and
an extensive discussion is presented about the advantages and shortcomings of the
previous evaluation approaches.Olmo and Gaudiosotried to project the presentation
and calculation of the RS separately [267]. Cremonesi and Lentinihave suggested the
procedure to evaluate collaborative filtering (CF) based recommender systems
R(S)[268]. They use 7 evaluation metrics to evaluate the system.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
119 | P a g e
The recommender system evaluation usually has three different approaches stated in
literature; offline evaluation, online evaluation and user studies [1], [10]. The author
[269] argues that online experiment has very difficult to be evaluated as it keeps on
generating data. They propose a prequential evaluation protocol that is useful for both
online and offline evaluation.
As far as evaluation metric are concerned, Mean Average Precision (MAP) is
supposed to be defacto standard, also Mean Reciprocal Rank (MRR) is considered
best for measuring the recommendation for top ranked products [38], [40]. In [39],
[270] authors discussed the idea of finding the errors in recommendation using false
positive and false negative and, illustrated how these help in finding the accuracy of
the system.
In the previous work [39], [40] authors have proposed evaluation process but no
criteria of preference is assured. They have quantified the user feedback and selected
all the items which have more than zero score, however, all the items for which user
visits to check the review, the value will come out to be more than zero, although it is
never guaranteed what have been visited to see is also the preference of a user. Here,
criteria of preference are defined mathematically, and it is suggested about an item
whether it can be treated as preferred product or not?
A survey based on more than 28,000 Internet users across 56 countries by Nielsen in
2012 suggests that online customer reviews are the second most trusted source of
brand information [271]. However, the biasedness and casual-review remain a
concern for the online reviews [29]. In various evaluation studies [267], [51] the
integrity of the user feedback is not well explored and the emphasis on the
consequences of fake, casual or biased feedback are not discussed. In this paper, a
check for biasedness and casualness is proposed by measuring user‟s sincerity. The
approach empowers the feedback, which in turn is the base for evaluation study.
6.3 Evaluation Metrics
Deciding how wella recommender system performs, depends upon the task it needs to
achieve. There are no absolute guidelines for designing recommender systems,
instead, it has to be designed in such a way that it may satisfy the users and meet their
needs. The users‟ needs vary according to the situation and as per their requirements.
Thus, the evaluation of such systems is purely relative rather than absolute.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
120 | P a g e
In the literature, prediction accuracy is the most considered factor of a
recommender system [51]. It refers to the accuracy of the recommender system in
predicting the items to the users. It is pre-assumed that users would prefer the systems
that have predicted the items for them more accurately. The prediction can be
classified in the following categories.
Accuracy of ratings predictions
Accuracy of usage predictions
Accuracy of rankings of items
There are different evaluations metrics have been used for deciding these accuracies.
Few of the most frequently used metrics are;
i. P@10
ii. FPR@10
iii. FNR@10
iv. Mean Average Precision (MAP)
v. Mean Absolute Error (MAE)
vi. Mean Reciprocal Rank (MRR)
vii. Root Mean Square Error (RMSE)
viii. Spearman rank Correlation Coefficient
ix. Modified Spearman rank Correlation Coefficient
The calculation method of the above metrics and their details are illustrated below.
6.3.1 P@10
We define the precision at top-k positions as P@k and it is given as;
' ' ‟ @
Number of products recommended by system ranking in top k positions that are also endorsed in user s raP k
k
nking
For different value of „k‟ we may obtain different P@k. for k=10, we define the
precision at top-10 positions as P@10 and it is given as;
@
'10' 10
1
‟
0
Number of products recommended by system ranking in top positions that are also endorsed in user s rankingP
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
121 | P a g e
6.3.2 FPR@10
We denote false positive rate for top 10 positions as FPR@10. We define FPR@10 as
follows:
'10 ' @10
1
0
Number of products recommended in top position but not preferred by customF R
erP
The “false positive” point out a situation in which recommended products are not
preferred by the customers. This situation is considered as the worst, as the customers
get irritated and never go for any further buying.
6.3.3 FNR@10
The false negative error refers to a situation when the preferred items customers are
missing in the recommendation. We denote false negative rate for top 10 positions as
FNR@10. We define FNR@10 as follows:
( '10' @10
10
Number of products missing in recommendation but preferred by customer in top positiFN
onR
6.3.4 Mean Average Precision
Mean Average Precision (MAP) is given by;
1
1( ) --------------------------------------------------- (6.1)
n
i
i
MAP P Un
P(Ui) is the precision of ith
user, and „n‟ is the number of users.
6.3.5 Mean Absolute Error
The Mean Absolute Error (MAE) is measured to know the closeness of the outcomes
with actual results. It is given by;
i
1
1x - x ------------------------------------------------ (6.2)
n
i
MAEn
xi and x are the outcomes and actual values respectively, and n is total number of
observations.
6.3.6 Mean Reciprocal Rank
In the ranking of the products, based on comprehensive approach, we find the position
of the product of an item which is ranked first in the system ranking. Let „r‟ denotes
its ranked-position in comprehensive ranking. We give Reciprocal Rank (RR) as:
RR = 1
r
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
122 | P a g e
Mean Reciprocal Rank (MRR) of all the items for their respective first ranked product
is given by;
1
1 -------------------------------- (6.3)
n
i
i
MRR RRn
Where „n‟ is the total number of different items and „i‟ denotes item‟s sequence.
Mean Reciprocal Rank (MRR) gives the degree of relevance of a particular product in
the eye of a customer. If all the first ranked product of different items in system
ranking is also ranked first in comprehensive ranking, MRR will be 1, i.e. the best
case.
6.3.7 Root Mean Square Error
The root mean square error gives the error value of the data. It is given as;
2
1
1( ) --------------------------------- (6.4)
n
i i
i
RMSE Y yn
Yi and yi are the actual ranking and outcomes of the ranking by experiments. In the
above equation, „Yi „in the equation denotes expert‟s recommendation, and „yi „is
used for indicating prediction by system.
The less the value of RMSE, the most accurate is the ranking.
6.3.8 Spearman rank Correlation Coefficient
Spearman‟s Rank correlation coefficient is used to identify and test the strength of a
relationship between two sets of data.
The Spearman rank correlation coefficient „ 'r ‟ is mathematically given by:
2
2' 1 6
( 1)
dr
N N
------------------------------ (6.5)
Where d is the difference in statistical rank of corresponding variables, and is an
approximation to the exact correlation coefficient.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
123 | P a g e
6.3.9 Modified Spearman rank Correlation Coefficient
The researchers have used spearman rank correlation coefficient for the measurement
of similarities between different rankings [272]. The problem with this coefficient is
its inability of producing correct correlation for partial list. Beg [273] have suggested
modified version of this coefficient. The formal definition of the modified spearman
rank correlation coefficient is given as-
“If full list is given as [1, 2,…, n] and let the partial list be given as [v1, v2,…,vm].
Then Without loss of generality, Modified Spearman rank order correlation
coefficient (rs׳) between these two rankings is given as follows”;
------------------------------ (6.6)
6.4 Evaluation based on Experts‟ Ranking using Explicit Feedback
The evaluation of recommender systems (RS) is very important in deciding the
procedure and parameters to be considered while designing RS. A quality evaluation
mitigates the issues encountered in recommendation and reinforces the
recommendation by placing the products before the users that matches to their
preferences. A detail discussion for the evaluation studies has been stated in section
6.2 and related evaluation metrics are mentioned in section 6.3.
The present study tries to help the academicians in deciding the books for the
syllabus of university graduates of India. On the one hand the proposed study helps
the university experts and administrators while designing the course curriculum for
their graduate students, and on the other hand it lays a platform for the students to
explore the challenging books for their studies. Thus, assessing whether the
recommendation approach performs well and would fulfill the requirement has great
significance. Therefore, we have evaluated the adopted procedure with the help of
experts. The term „expert‟ refers to the specialist of computer science. These
specialists are selected from different locations on the globe with only consideration
in common is their direct or indirect connection with the Indian education system
which involves computer science to an extent.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
124 | P a g e
The list of recommended books by the book recommendation approaches (BRA)
which are discussed in previous chapters 3, 4 and 5 are provided to the specialists.
The specialists are computer science graduates, researchers, corporate officials and
academicians. They are Indian and foreigners both, but have link with computer
science education in India. It is described earlier in Chapter 3 that a total of 10
different courses are taken. 10 different specialists for each course are approached.
They are searched from their profile, university data base and Google scholar.
Especially the faculty members of Computers Science Department of central
universities of India, research scholar at Indian universities in the Department of
Computer Science and IT industries engineers are included. The experts from India,
Iraq, Iran, USA, KSA, Jordan, Yemen, UK and Australia are approached. The
specialists from abroad are the researchers and IT specialist working in India or the
Indian computer science graduates working in those countries. The ranked books
from the BRA are compared with the ranking of books provided by experts. On the
basis of several evaluation measures the recommendation approaches are evaluated. A
block diagram for illustrating the evaluation scheme is shown in Figure 6.1.
The block labeled with „prescribed books by top ranked universities‟ shows the list
of books prescribed in the universities‟ curriculum. These books are used by all the
different recommendation approaches (BRA). The details of all these approaches have
been extensively discussed in respective sections of Chapter 3 and Chapter 4. The
final rankings of different approaches have been stored and these different lists of
ranked books are considered as the final recommendation by the book recommender
respectively.
As shown in the above figure, the list of books is given to experts. Since, we have
10 different courses containing varying books which has a total of 158 in number. The
books from each course are presented to their corresponding experts. Let us consider
books on “operating systems” (OS). There are 15 different books of OS. These 15
books of OS are given to experts and their explicit feedbacks for the ranking of these
books are recorded.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
125 | P a g e
Figure 6.1: Block diagram for Evaluation of Book Recommendation Approaches
The final recommendations of books by each recommender approach are compared
with the ranking given by experts. On the basis of evaluation measures which includes
Mean Reciprocal Rank (MRR), P@k, Mean Average Precision (MAP), FPR@k (false
positive rate), FNR@k (false negative rate), Mean Absolute Error (MAE), Root Mean
Square Error (RMSE) and Modified Spearman rank correlation coefficient, the
recommendation are evaluated.
6.4.1 Evaluation Results based on Different Evaluation Metrics
The different evaluation metrics have been discussed in 6.3. These metrics are
frequently used to evaluate the performance of recommender systems. We have used
eight (8) metrics from the above discussed sections for the purpose of evaluation of
our proposed recommender system based on the experts ranking. The different
approaches for the recommender system have been compared and the results from
various aspects have been shown and discussed.
The all parameters which are included in the evaluation process have been
discussed one by one to show their values for each technique. Hence, the relative
comparison of the recommendation approaches for the respective parameters are
presented.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
126 | P a g e
6.4.1.1Evaluation Results using Explicit Feedback based on Root Mean Square Error
The Root Mean Square Error (RMSE) of ranking of books provided in Chapter 3,
Chapter 4 and Chapter 5 are compared with the experts‟ ranking. For all 10 different
courses the values of RMSE for OMT, PAS, OWA with different quantifiers and
ORWA is shown in the Table 6.1. The RMSE is obtained using equation 6.1.
The RMSE for the book „Theory of Computation‟ (TOC) is least for the Opinion
Mining Technique (OMT). It comes out to be 0.645 which is the minimum value and
OWA (at least half) has performed well next to OMT as RMSE for it has obtained as
0.936. The maximum RMSE, i.e. the worst performance has been recorded by OWA
(as many as possible). However the value is still not too much and considerably low.
This variation indicates that the recommendation made by the techniques based on
user‟s opinion has least error while compared with experts ranking for the books on
TOC. A similar trend has been noticed for books on Data Base (DB). This again
emphasizes that the techniques used for recommendation based on their online
opinion has the most similarity to the experts‟ rankings. It suggests that the adopted
method is most accurate and the experts ranking has least difference with readers‟
opinions. Surprisingly, Ordered Ranked Weighted Aggregation (ORWA) has the
minimum RMSE for books on Compiler Design (CD) and its value is same as RMSE
obtained for OWA (at least half). The reason behind the similar value of ORWA and
OWA (at least half) is the weight assignment methodology of ORWA.
Table 6.1: Root Mean Square Error of all books by different approaches
Courses
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
TOC 0.97 0.96 0.94 1.14 1.01 0.65
DB 0.86 0.86 0.81 1.11 0.91 0.74
CD 0.91 0.73 0.73 1.06 0.91 0.82
OS 0.86 0.87 0.88 1.32 1.06 0.84
DS 0.95 0.90 0.91 1.11 1.17 0.96
AI 2.22 2.16 2.34 2.46 2.23 2.00
CN 1.40 1.46 1.36 1.74 1.80 1.17
SE 1.08 1.00 1.10 1.48 1.61 1.17
DM 0.99 0.80 1.25 1.61 1.37 1.08
CG 2.14 1.52 1.74 2.58 2.59 0.29
Average
RMSE for
all books
1.24 1.13 1.21 1.56 1.47 0.97
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
127 | P a g e
Since, while assigning weights using ORWA, the ranking of the rankers are taken
into considerations. The OWA (at least half) also considers the upper half of the best
ranked institution. Therefore, in most of the cases these two methods have similar
value of the metric. The maximum difference between the values of ORWA and
OWA (at least half) is observed for books on „Computer Network‟ (CN) and „Discrete
mathematics‟ (DM). The possible reason behind this could be the variation in the
recommendation of CN and DM books by universities for their students. Due to
varying books prescribed in the syllabus by universities less number of similar books
are found which in turn gives the difference in values of RMSE for ORWA and OWA
(at least half).
The root mean square value of different books is found to be least by OMT in most
of the cases. And the book on AI has most error by all the methods. The OMT has
least error in 6 out of 10 cases whereas ORWA has 4 times better performance than
other techniques. OWA (at least half) has the second best performance in two cases
and Positional Aggregation Scheme in three cases respectively. However, OWA (as
many as possible) has the maximum RMSE in most of the cases. The major difference
in the values of respective techniques is observed for books on „Computer Graphics‟
(CG). The experts‟ recommendation and user‟s opinions have least root mean square
error. It simply implies the techniques adopted for opinion extraction and
recommendation of products is up to the mark and it shows the strong similarities
between experts ranking and user‟s reviews. Also, the results indicate that the
interpretation of opinions are successfully framed that it has high similarities with the
recommendation of experts.
The average values of the RMSE for all the books are represented pictorially in
Figure 6.2. The result gives the clear indication of OMT being outperformer for these
parameters as it holds the least error. The average RMSE is maximum for OWA (as
many as possible) and minimum for OMT. ORWA is the second best performer as far
as RMSE is concerned. Unlike ORWA, OWA (at least half) has not obtained equally
least RMSE. Since it includes the half of the rankers, i.e. universities in
recommendation process and there are the books recommended by least ranked
universities also which in turn down the ranking of this method. Thus the results
shows the significance of the authoritative recommendation as the results of these
techniques i.e., ORWA, OWA with all quantifiers and PAS are authoritative
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
128 | P a g e
recommendations considers universities authorities suggestion while making
recommendation for students their best books.
On observing the RMSE for different techniques, one interesting thing that can
easily be seen is that the maximum RMSE which has been attained by OMT is 2.0.
Thus, OMT not only has least error for most of the books but also shows the
minimum error while considering the maximum value of RMSE for each of the
techniques.
Figure 6.2: Average Root Mean Square Error for all techniques
6.4.1.2 Evaluation Results using Explicit Feedback based on Mean Absolute Error
The Mean Absolute Error (MAE) of ranking of books provided in Chapter 3,
chapter 4 and Chapter 5 are compared with the experts‟ ranking. For all 10 different
courses the values of MAE for OMT, PAS, OWA with different quantifiers and
ORWA is shown in the Table 6.2.
The MAE for the book „Computer Network‟ (CN) is least for the ORWA. It comes
out to be 4.54. The maximum MAE, i.e. the worst performance has been recorded by
OWA (as many as possible). This variation indicates that the recommendation made
by the techniques based on ordered ranked weighted aggregation which is applied
over authoritative recommendations has least error when compared with experts
ranking for the books on CN. Unlike RMSE, the results of MAE have lowest for all
books using ORWA. The consistent performance has been noticed for related
approaches, hence, PAS and OWA (at least half) has also performed well.
The results have almost same trend in every technique for all the books. Though,
the differences between ORWA and OMT values of MAE have variations. For books
TOC, DB, AI and CD there are minor differences whereas MAE for other books has
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
129 | P a g e
slightly more differences. It suggests that the adopted method of aggregation for the
authoritative recommendations has the best performance as far as measurement of
MAE is concerned. Still, OWA (as many as possible) and OWA (most)
recommendations lag behind. The reason for this is basically the concept that these
two methods consider lower half of recommendation while performing aggregation.
Hence, instead of high ranked rankers‟ recommendation it includes the lower ranked
recommendations. Also, Ordered Ranked Weighted Aggregation (ORWA) has the
minimum MAE for books on Compiler Design (CD) and its value is same as MAE
obtained for OWA (at least half). The reason behind the similar value of ORWA and
OWA (at least half) is the weight assignment methodology of ORWA.
Table 6.2:Mean Absolute Error of all books for different approaches
Courses
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
CN 5.91 4.54 4.25 6.84 6.06 5.52
DM 3.83 2.81 3.94 6.22 4.89 4.338
OS 2.73 2.76 2.77 4.19 3.35 3.16
TOC 3.05 2.91 2.89 3.51 3.13 3.098
SE 4.41 3.23 3.41 6.23 5.59 4.574
DBMS 2.72 2.73 2.57 3.5 2.89 2.882
DS 2.98 2.81 2.8 3.54 3.71 3.168
AI 6.7 6.45 6.99 7.21 6.75 6.82
CD 2.88 2.3 2.3 3.36 2.88 2.744
CG 4.41 3.23 3.41 6.23 5.59 4.574
Average
MAE for all
the books
3.962 3.377 3.533 5.083 4.484 4.0878
The MAE value of different books is found to be least by ORWA in all the cases.
Like RMSE, the book on AI has most error by all the methods. However, the least
Mae is obtained for CD in all the cases. As total number of books on CD is only 10
therefore the overall recommendations have fewer variations. The major difference in
the values of respective techniques is observed for books on „Computer Graphics‟
(CG).
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
130 | P a g e
Figure 6.3: Average of Mean Absolute Error of all the books for different quantifiers
6.4.1.3 Evaluation Results using Explicit Feedback based onP@10
The P@10 gives the value of preciseness in recommendation that how many books
which have been recommended by the Experts in top 10 positions is also
recommended by Book Recommendation Approaches (BRA). This measure tells how
accurate the adopted recommender technique is and how accurate is the
recommendation made by these systems.
The values of P@10 for the respective books by using all different
recommendation techniques are given in Table 6.3. The ORWA has maximum P@10
for the books of four courses whereas OMT has higher values of P@10 for five
courses. Thus the recommendation which matches most to the expert‟s ranking is
made by OMT, ORWA and OWA (at least half); hence these are the best performers
as far as P@10 is concerned. Like ORWA, OWA (at least half) has obtained equally
good values of P@10 and almost for each book the values are same for both the
method.
For books on „Compiler Design‟ (CD), P@10 is 1 for all the techniques. This is
again because of the same reason that CD has a total of10 books only, and experts
have ranked them in order. This makes all the books to anyhow fall in the
recommendation list of experts. Similarly, all 10 books are also ordered from best to
least by each technique, which in turn allows all 10 books to be recommended.
Therefore, the method could have been consider true adoptable for the large number
of books, where top 10 books out of the large number of collection of books would
mean really a better scrutinized results.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
131 | P a g e
Thus, the results show the significance of the authoritative recommendation as the
results of these techniques i.e., ORWA, OWA with all quantifiers and PAS are
authoritative recommendations, consider universities authorities suggestion while
making recommendation for students their best books.
By observing the average P@10 for different techniques which is shown in Figure
6.4, one interesting thing that can easily be seen is that the average values of P@10
for different approaches has insignificant differences and are very close to each other.
ORWA and OWA (at least half) have [email protected]. The PAS and OWA (most) has 0.72
and 0.70 respectively. The maximum value of average P@10 for all the books using
each technique is 0.78 which is acquired by Opinion Mining Technique (OMT).
Table 6.3: P@10 for all approaches
Courses
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As many
as possible)
OWA
(Most) OMT
CN 0.53 0.67 0.65 0.45 0.57 0.73
DM 0.75 0.77 0.77 0.45 0.65 0.68
OS 0.74 0.72 0.72 0.7 0.74 0.78
TOC 0.92 0.92 0.92 0.91 0.9 0.91
SE 0.6 0.76 0.76 0.44 0.56 0.75
DBMS 0.84 0.81 0.81 0.78 0.81 0.82
DS 0.7 0.74 0.74 0.69 0.69 0.78
AI 0.53 0.54 0.53 0.5 0.53 0.61
CD 1 1 1 1 1 1
CG 0.6 0.76 0.76 0.44 0.56 0.75
Average
P@10 for all
the books
0.721 0.769 0.766 0.636 0.701 0.781
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
132 | P a g e
Figure 6.4:Average P@10 for all books
Thus, again it is evident that the experts‟ suggestion and users‟ opinion coincide.
On the one hand it can be interpreted as the proficiency of the technique adopted
which have been designed in such a way that users sentiments are converted to rank
using feature based extraction and evaluation of reviews, on the other hand it is
supports and proves the philosophy of selecting experts‟ ranking as base for the
evaluation of the systems.
6.4.1.4 Evaluation Results using Explicit Feedback based on Mean Average Precision
The average of MAP for the respective techniques is presented in Table 6.4. The
OMT has performed relatively well when we measure P@10. In the same way, Mean
Average Precision (MAP) comes out to be highest for OMT. MAP for Opinion
mining technique (OMT) which is discussed in Chapter 5 is 0.6397 and MAP for
ORWA is 0.5571. The difference in the value is because of the variation in the
precision for different „k‟ in the value of P@k, at 10th
position the value of precision
may have higher values whereas for other values of k, it might have significant
changes. The MAP incorporates precision for each position and hence gives a more
holistic representation of precision.
Table 6.4: Mean Average Precision of different approaches.
PAS ORWA
OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
MAP for all courses
altogether 0.5257 0.5733 0.5571 0.4011 0.4593 0.6397
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
133 | P a g e
Figure 6.5: Mean Average Precision of different approaches
6.4.1.5 Evaluation Results using Explicit Feedback based on FPR@10
The FPR@10 for all the techniques is given in Table 6.5. The FPR@10 gives the
value of impreciseness in recommendation that how many books which have been
recommended by the Experts in top 10 positions is not recommended by Book
Recommendation Approaches (BRA). This measure tells how accurate the adopted
recommender technique is and what degree of false positive appears in the
recommendation made by these systems.
The value of FPR@10 for the respective books by using all different
recommendation techniques is least for ORWA for the books of four courses whereas
OMT has minimum error values for five courses. Thus the recommendation which
matches most to the expert‟s ranking is made by OMT and ORWA. Hence, these are
the best performers as far as false positive error for top 10 positions, i.e. FPR@10 is
concerned. Like ORWA, OWA (at least half) has obtained equally good values and
almost for each book the values are same for both the method.
False positive value of different books is found to be least by OMT in most of the
cases. And the book on AI has most error by all the methods. The OMT has least error
in 5 out of 10 cases whereas ORWA has 4 times better performance than other
techniques. OWA (at least half) has the performance similar to ORWA and Positional
Aggregation Scheme relative performance for books on „data base‟ (DB) is better than
any other technique. However, on analyzing the average FPR@10, OWA (as many as
possible) has the maximum FPR@10 in most of the cases. The major difference in the
values of respective techniques is observed for books on „Computer Graphics‟ (CG)
and „Discrete Mathematics‟ (DM). Also, the results indicate that the interpretation of
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
134 | P a g e
opinions are successfully framed that it has high similarities with the recommendation
of experts.
Table 6.5: FPR@10 for all techniques.
Courses
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
CN 0.47 0.33 0.35 0.55 0.43 0.27
DM 0.25 0.23 0.23 0.55 0.35 0.32
OS 0.26 0.28 0.28 0.3 0.26 0.22
TOC 0.08 0.08 0.08 0.09 0.1 0.09
SE 0.4 0.24 0.24 0.56 0.44 0.25
DBMS 0.16 0.19 0.19 0.22 0.19 0.18
DS 0.3 0.26 0.26 0.31 0.31 0.22
AI 0.47 0.46 0.47 0.5 0.47 0.39
CD 0 0 0 0 0 0
CG 0.4 0.24 0.24 0.56 0.44 0.25
Average
FPR@10
for all the
books 0.279 0.231 0.234 0.364 0.299 0.219
Figure 6.6: Average FPR@10 for all books using different book recommender approaches
6.4.1.6 Evaluation Results using Explicit Feedback based on FNR@10
The FNR@10 for all the techniques is given in Table 6.6. The FNR@10 gives the
value of impreciseness in recommendation that how many books which have been
recommended by the Experts in top 10 positions is not recommended by Book
Recommendation Approaches (BRA). This measure tells how accurate the adopted
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
135 | P a g e
recommender technique is and what degree of false negative appears in the
recommendation made by these systems.
The value of FNR@10 for the respective books by using all different
recommendation techniques is least for ORWA for the books of four courses whereas
OMT has minimum error values for five courses. Thus the recommendation which
matches most to the expert‟s ranking is made by OMT and ORWA. Hence, these are
the best performers as far as false negative error for top 10 positions, i.e. FNR@10 is
concerned. Like ORWA, OWA (at least half) has obtained equally good values and
almost for each book the values are same for both the method.
False negative value of different books is found to be least by OMT in most of the
cases. And the book on AI has most error by all the methods. The OMT has least error
in 5 out of 10 cases whereas ORWA has 4 times better performance than other
techniques. OWA (at least half) has the performance similar to ORWA and Positional
Aggregation Scheme relative performance for books on „data base‟ (DB) is better than
any other technique
Table 6.6: FNR@10 of allbooks
Courses
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA (Most) OMT
CN 0.47 0.33 0.35 0.55 0.43 0.27
DM 0.25 0.23 0.23 0.55 0.35 0.32
OS 0.26 0.28 0.28 0.3 0.26 0.22
TOC 0.08 0.08 0.08 0.09 0.1 0.09
SE 0.4 0.24 0.24 0.56 0.44 0.25
DBMS 0.16 0.19 0.19 0.22 0.19 0.18
DS 0.3 0.26 0.26 0.31 0.31 0.22
AI 0.47 0.46 0.47 0.5 0.47 0.39
CD 0 0 0 0 0 0
CG 0.4 0.24 0.24 0.56 0.44 0.25
Average
FNR@10 for
all the books 0.279 0.231 0.234 0.364 0.299 0.219
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
136 | P a g e
Figure 6.7: Average FPR@10 for all books using different book recommender approaches
However, on analyzing the average FNR@10, OWA (as many as possible) has the
maximum FNR@10 in most of the cases. The major difference in the values of
respective techniques is observed for books on „Computer Graphics‟ (CG) and
„Discrete Mathematics‟ (DM). Also, the results indicate that the interpretation of
opinions are successfully framed that it has high similarities with the recommendation
of experts.
6.4.1.7 Evaluation Results using Explicit Feedback based on Modified Spearman Rank
Correlation Coefficient
The researchers have used spearman rank correlation coefficient for the
measurement of similarities between different rankings [272]. The problem with this
coefficient is its inability of producing correct correlation for partial list. Beg [273]
have suggested modified version of this coefficient. The modified rank correlation
gives the degree of similarities between the two different rankings. Here, we have
proposed 6 different book recommendation approaches; hence it leads to 6 different
rankings of books. These 6 rankings are compared with the ranking of books collected
from experts and considered as standard rankings. The modified spearman correlation
coefficient are expressed in Table 6.7
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
137 | P a g e
Table 6.7: Modified Spearman Rank Correlation Coefficient by different approaches
Courses
Recommendation Approaches
PAS ORWA
OWA
(At least
half)
OWA (As
many as
possible)
OWA
(Most) OMT
TOC 0.87 0.88 0.88 0.85 0.87 0.93
DB 0.91 0.92 0.92 0.88 0.91 0.93
CD 0.87 0.90 0.90 0.83 0.86 0.88
OS 0.93 0.93 0.93 0.87 0.90 0.93
dS 0.93 0.93 0.92 0.89 0.89 0.93
AI 0.82 0.83 0.86 0.80 0.82 0.82
CN 0.88 0.88 0.88 0.81 0.82 0.92
SE 0.92 0.91 0.90 0.87 0.84 0.90
DS 0.93 0.93 0.88 0.86 0.89 0.92
CG 0.80 0.89 0.87 0.76 0.76 0.98
Average MSRCC 0.89 0.90 0.89 0.84 0.86 0.91
Figure 6.8: Average of Modified Spearman rank correlation coefficient
The minimum value of average spearman rank correlation coefficient is 0.842
which have been observed while calculating for OWA (as many as possible). The
individual score of average correlation for each book is very close for all techniques.
An insignificant difference can be observed, however, for books on „Computer
Graphics‟ (CG) maximum difference in the correlation values for respective
techniques is recorded. The highest correlation with the experts‟ rankings is recorded
for OMT. Average Correlation for all the books by using OMT is 0.91 which is the
maximum average correlation. If considering average correlation for books
separately, the most correlated observation is seen for books on CG with OMT which
is 0.97.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
138 | P a g e
The highly correlated values of each techniques with experts‟ rankings has clear
indications of all the adopted techniques being close to the what can help the users
and can be used to satisfy the users‟ needs and provide them with accurate
recommendations.
6.4.1.8 Evaluation Results using Explicit Feedback based on Mean Reciprocal Rank
The average MRR for all the techniques is presented in Table 6.8 and shown in
Figure 6.9. The average MRR for ORWA is nearly 0.61. It means 61% times the 1st
ranked books recommended by ORWA is also ranked 1 by experts in their rankings.
Basically MRR tries to let users aware of the best products and recommend it along
with a list of complete ranked products.
Table 6.8: Mean Reciprocal Rank of all techniques for different Courses
Courses
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
CN 0.50 0.58 0.55 0.46 0.25 0.49
DS 0.83 0.84 0.87 0.15 0.30 0.86
OS 0.85 0.88 0.88 0.42 0.44 0.88
TOC 0.59 0.57 0.58 0.57 0.60 0.66
SE 0.65 0.73 0.61 0.55 0.54 0.59
DBMS 0.62 0.63 0.78 0.51 0.53 0.58
dS 0.58 0.58 0.49 0.23 0.25 0.45
AI 0.30 0.27 0.25 0.33 0.29 0.26
CD 0.56 0.56 0.56 0.56 0.58 0.57
CG 0.65 0.73 0.61 0.55 0.54 0.58
Average
MRR for
all books
0.613 0.637 0.618 0.433 0.432 0.592
The MRR for PAS and OWA (at least half) is relatively more close to the experts‟
suggestion than OMT. The OMT has average MRR 0.59 whereas MRR for PAS and
OWA (at least half) are 0.61, approximately same.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
139 | P a g e
Figure 6.9: Average Mean Reciprocal Rank of all the books for different techniques
6.4.2 Comprehensive Evaluation Measure
The different parameters have been used in this section and have discussed in details.
To know which of the adopted recommendation approach has performed better, we
have suggested a comprehensive evaluation measure which aggregates all the above
evaluation metrics uniformly. The two different types of metrics have been used in
this chapter. One metric finds the error, which we call as fallacy metric and another
technique measures precision, which can be termed as veracity metric. The veracity
measures should be high and fallacy measure should be low for a good recommender
system. These two metrics have been shown in the Table 6.9 and Table 6.10, and their
aggregated comprehensive value is shown in Table 6.11.
The comprehensive evaluation measure (CEM) is calculated as;
( values of veracity metrics)+( values of fallacy metrics)Comprehensive Evaluation Measure =
metrics
sum of sum of
number of----(6.7)
Table 6.9: Final values of parameters used to find error
Metrics
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
RMSE 1.23 1.12 1.2 1.56 1.46 0.97
FPR@10 0.28 0.23 0.23 0.36 0.3 0.22
FNR@10 0.28 0.23 0.23 0.36 0.3 0.22
MAE 3.96 3.38 3.53 5.08 4.48 3.29
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
140 | P a g e
Table 6.10: Final values of parameters used to find precisions and correlation
Metrics
Recommendation Approaches
PAS ORWA OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
P@10 0.72 0.77 0.77 0.64 0.7 0.78
MAP 0.55 0.61 0.6 0.42 0.47 0.62
Correlation 0.34 0.41 0.39 0.14 0.14 0.43
MRR 0.61 0.63 0.62 0.43 0.43 0.59
Table 6.11: Comprehensive evaluation measure
PAS ORWA
OWA (At
least half)
OWA (As
many as
possible)
OWA
(Most) OMT
Comprehensive
Evaluation
Measure
1.304 1.538 1.524 1.003 1.164 1.606
The Comprehensive Evaluation Measure (CEM) gives an aggregated interpretation
of all the techniques used. With all different metrics and distinguished
recommendation approaches, a final recommendation by the proposed evaluation
approach is feature extraction based opinion mining technique (OMT). CEM for OMT
is 1.606 which is marginally ahead from OWA (as many as possible) and OWA
(most).
6.5 Evaluation based on Implicit User Feedback
In this section, we have evaluated a recommender system and suggest a
comprehensive approach for the evaluation of a recommender system so that user
satisfaction can be assured. We perform user studies to evaluate recommender system.
We present the different recommended items and links to their reviews from a
recommender system proposed in [6] to users. Implicit user feedbacks were taken that
captures the behavior of the users over the recommended items of the recommender
system. We quantify the feedback and associate a score to each product from each
user. The sincerity of users is measured quantitatively and only feedbacks from
sincere users are considered for evaluation process. We put a threshold to classify the
preference of a user and only preferred items for users are stored. It gives ranking of
preferred products for all users. We get user‟s aggregated ranking of products for all
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
141 | P a g e
users by using rank aggregation algorithm. We call „system ranking‟ to ranking
presented by the system which is being evaluated. The system ranking is evaluated on
the basis of aggregated ranking of products. The system is evaluated using several
evaluation metrics. Also a relative comparison with related approach is done.
We can summarize our main contributions for Implicit evaluation of RS as follows:
I) Though quantification of user feedback has been done[274] but there is no
sincerity measure for the users. We have introduced user‟s sincerity
measure that strengthens the procedure and hence, make the evaluation
process robust.
II) Though sincerity measure improves the fairness in evaluation process, we
set criteria of preference for the products to classify the user‟s preferred
products. We put a threshold value for the product importance score that
invigorated the criteria of preference.
III) Accuracy alone may not decide the quality of a recommender system,
user‟s interaction and satisfaction is important. We suggested an approach
that interacts with users and took feedback from them to evaluate the
recommender system.
IV) The proposed approach is fulfilling two aspects simultaneously. First, it
gives a comprehend methodology to evaluate a recommender system that
can be generalized to any product and any marketing portal. Second, the
methodology is implemented to evaluate a recommender system and tells
how that system performance is.
V) Biasedness and casual-reviews are major concern in user feedback. We
have explicitly defined and described a procedure to trace the insincerity in
exploiting user feedback.
6.6 Architecture for Evaluation Scheme based on Implicit Feedback
We give architecture (Figure 6.10) for evaluation method of recommender system.
The evaluation procedure is basically two-step process. First, step is to lay down a
recommendation approach and second is to compare the recommended items with that
of which is recommended by the system under evaluation. The recommendation
approach constitutes of several steps. Initially, final recommendation given by a
recommender system [243] is presented before users. The recommender system
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
142 | P a g e
recommends top 10 products of 5 different items. The list of all ranked products and
link to reviews of their respective products are given to 10 different users. All the
users are either graduate students or professionals familiar with the use of Internet.
Figure 6.10: Block diagram of implicit User Feedback based Evaluation of Recommender
Systems
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
143 | P a g e
We have taken implicit user feedback (see section 6.6.1) for these reviews and
quantify the feedbacks (see section 6.6.2). The quantification of the feedbacks give us
product importance score (PIS) for the products by respective users. Outlieranalysis is
performed to check the sincerity of the users (see section 6.6.3). Outlier analysis gives
user‟s sincerity measure. The sincerity measure of the users strengthens the ranking
process as it allows inclusion of only sincere and substantial users in the
ranking.Athreshold value for PIS to set criteria of preference (see section 6.6.4) is
suggested. Onlythose products that fall under the criteria of preference are considered,
i.e. products for which PIS are higher than the threshold value. By sorting these
products, we get different ranking of each product for each user.We apply a rank
aggregation algorithm to aggregate the ranking of all users to get a final ranking. This
final ranking is the aggregated ranking of products by the users. We have evaluated
the system ranking on the basis of aggregated user‟s ranking of products.
6.6.1 Vector Component of User feedback
To observe the user behavior, we take implicit user feedback in vector form. We use
five vectors namely E, P, S, T and V to observe the user‟s behavior, thus we call the
feedback as vector feedback. The five vectors are explained below.
(a) V: it is the sequence „V‟ of visiting the review site of the product by a user.
Let there be „n‟ number of products whose reviews are available for „m‟ users
and ith user (i≤m) visits a review site of jth product (j≤n) which is the kth site
visited by the user. We assign vij = k, where k≤ n. A snippet about the product
is attached with the links. Thus, users would visit the page which they find
more close to their choices. Hence, the sequence in which a page is visited
shows the significance of product in the eyes of a user.
(b) T: The time duration that ith
user gives to reconnoitre the reviews for the
product „j‟ is denoted by Tij. The value of T is assigned „0‟for a product whose
review is not visited by the user.
(c) We use Boolean „P‟ to denote whether or not a user prints the review of a
product. The Boolean vector P is denoted by „Pij‟. Where „i‟denote ith
user and
„j‟ denote jth
product.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
144 | P a g e
(d) We use Boolean „S‟ to denote whether or not a user saves the review of a
product. The Boolean vector S is denoted by „Sij‟. Where „i‟denote ith
user and
„j‟ denote jth
product.
(e) We use Boolean „E‟ to denote whether or not a user e-mail the review of a
product. The Boolean vector E is denoted by „Eij‟. Where „i‟denote ith
user and
„j‟ denote jth
product.
The idea behind compiling this feedback is based on the assumption that an intelligent
user is likely to visit the more appealing product early in the system discovery
process. The snippet of the links would attract the user about the contents of the items;
hence, it is more probable that users would click for the links of items in the order of
their preference. Similarly, the time that a user invest in exploring a review, whether
or not the user saves it to their computer, and whether the user prints or e-mails it to
someone else reveals the degree of importance that a product holds for that particular
user[273].
Example 6.1: let 10 documents d1, d2... d10 are presented before a user, the user in
very first visit go to explore d3, and email the document to a friend. Spend 40 seconds
in traversing the documents. The user neither saves the document nor prints it. The
values of the vector component for d3 for the user would be, v=1, t=40 sec, e=1,
p=s=0.
6.6.2 User Feedback based Scoring of Products
To assess the user‟s behavior, we give a formula for the quantification of vector
feedback. The quantification will help in scoring the products‟ value and ranking
them accordingly [20]. We denote product‟s importance score (PIS) by ϕ. The PIS of
jth
product by ith
user is written as ϕij.
We give ϕij as:
ϕij = ij
1 1
{ (1/ )}n m
ij ij ij ij
i j
E P S T V
------------- (6.8); 0ijV
Where E, P and S are Boolean vectors acquire value 0 or 1 only. As described in
section 6.6.1, if a user prints, saves or emails a review to someone else, the Boolean
value will be set to 1 else 0 for respective Boolean measures. The importance of each
component may differ and would be identified if a weight is associated to these
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
145 | P a g e
according to their significance suggested by user‟s behaviours. The weighted sum is
given by equation 6.3.
ij
1 1
Sum = { (1/ )} (6.9)ij ij ij ij ij
n m
E P ij S ij T ij V ij
i j
Weighted W E W P W S W T W V
For the sake of simplicity in the procedure, we take all weights as 1. Hence, equation
6.3again reduces to equation 6.2. For calculation of time „T‟, we find that the average
lengths of the reviews are nearly about 400 bytes i.e. 400 characters. The reading
speed of a user is considered as 10 bytes/second. Thus, we classify different time
interval to calculate time „T‟. For ith user and jth product we have[273];
tij =0 if user do not visit the page.
tij =1 if time duration „t‟ spent in investigating review lies between 1 second to 39
second, i.e. 1≤t≤39
tij =2 if 40≤t≤79
tij =3 if 80≤t≤119
tij =4 if 120≤t≤159
tij =5 if t ≥160 seconds.
Finally we compute Tij as,
Tij= tij/tmax;
Here tmax =5;
If the user browses a review site for 45 seconds, we assign tij=2,
Tij=2/5;
Tij=2/5;
Tij =0.4;
To calculate ϕ for above case, we use Tij as 0.4.
A detail review with more number of characters, whether positive or negative, could
take longer time. In this case it seems the higher value of „T‟ does not reflect the
exactly high affinity of user for the reviewed item. However, we emphasize that a user
can only read a detail negative review of a highly desirable item. Thus, irrespective of
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
146 | P a g e
the polarity of the review, the product would be given more weights and must be
considered as a preferable one for the users.
We set the value of „V‟ in the sequence in which the review site is hit, i.e. if user „i‟
visits the jth product in very first click we set Vij = 1, this gives 1/ Vij = 1. If a user
does not visit a particular product‟s review, we set ϕ = 0.
As we use five vectors, we give normalized product‟s importance score (NPIS). We
denote NPIS of jth
product by ith
user as δij, and it is given by:
δij = ϕij / 5 --------------- (6.10)
We have calculated NPIS, „δ‟ of each product for each user. We have five different
items; Laptop, Head Phone, Smart Phone, Printer and Tablet. Each item has 10
different products. Here, in the Table 6.12, symbols L1, L2 to L10 represents 10
different ranked products of Laptops. These products are recommended by the
recommender system which is under study for evaluation. An illustration for the
calculation of NVS is summarized in the Table 6.12. The respective value of different
vector components of product L1, L2, etc. is calculated for user1 and calculation is
performed as mentioned in above equations. The δ of L1 is 0.92 i.e. the normalized
vector score of product L1 for user1 is 0.92. The value of all the components for
product L9 is 0; it shows that the review of the product is not visited by the user1 at
all.
Table 6.12: Illustration for the calculation of Normalized Products Importance Score „δ‟ for user
1.
Products E P S T T V 1/V ϕ δ
L1 1 1 1 3 0.6 1 1 4.6 0.92
L2 0 1 1 2 0.4 4 0.25 2.65 0.53
L3 0 1 1 2 0.4 2 0.5 2.90 0.58
L4 0 0 1 2 0.4 3 0.33 1.73 0.346
L5 1 0 0 3 0.6 5 0.2 1.80 0.36
L6 0 0 1 2 0.4 9 0.11 1.51 0.302
L7 0 1 0 1 0.2 6 0.16 1.36 0.272
L8 1 0 0 2 0.4 7 0.14 1.54 0.308
L9 0 0 0 0 0 0 0 0 0
L10 0 1 1 2 0.4 8 0.125 2.525 0.509
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
147 | P a g e
6.6.3 User‟s Sincerity Measure
Our approach to evaluate recommender systems is based on users‟ studies. We have
10 different users for each product. It is necessary to check whether the users are
sincere or not? As the sincerity of a user plays an important role in making any
decision based on the feedback taken from users. Any offhand feedback from
insincere users may lead to misguided conclusion. Therefore we measure user‟s
sincerity to know that user is sincere enough that the feedback from user can be
considered as a base in recommender system‟s evaluation process.
We find the correlation of all the users among themselves, represented as corr
(ui,uj), where i and j are users, for all the different products, and considered those
users as insincere for which we get a negative correlation. Finding out offhand
feedbacks strengthens the dataset and reduces the chances of data discrepancy. And it
gives a true measure of user‟s sincerity. The correlation values for laptop of all the
users are shown in Table 6.13, the user 5 has a negative correlation and we excluded
its feedback from our recommendation process.
Let, F is the function defining sincerity (s) of user „ui‟; we have: F(s,ui) = corr
(ui,uj) ; 1≤j≤10, j≠i
Thus, formally, user sincerity of a user „ui‟ is defined as;
≥ 0, Sincere user
F(s,ui)
< 0, Insincere user
A user‟s approach may vary with time and hence it is important to check the user‟s
sincerity for each item explicitly. The separate test of user‟s sincerity for each item
ensures true users‟ feedback of respective products of all the items. We find user‟s
sincerity for all the items in the same way as stated above for Laptop. The correlation
values of all the items, Head Phone, Smart Phone, Printer and Tablet has been shown
in Table 6.14 to Table 6.17, respectively. The user with negative correlation, i.e.
F(s,ui) < 0, is considered as insincere and represented in bold letters. The list of items
and corresponding users, whose offhand feedback is excluded from evaluation
process, is listed in Table 6.18.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
148 | P a g e
Table 6.13: Correlation values of different products of Laptop
User 1 user 2 user 3 user 4 user 5 user 6 user 7 user 8 user 9 user 10 Average
User 1 1.00 0.38 0.09 0.68 -0.04 0.53 0.79 0.72 0.76 0.65 0.56
user 2 0.38 1.00 0.43 0.85 -0.62 0.56 0.31 0.85 0.84 0.68 0.53
user 3 0.09 0.43 1.00 0.35 -0.70 0.28 0.00 0.35 0.34 0.35 0.25
user 4 0.68 0.85 0.35 1.00 -0.55 0.52 0.56 1.00 0.99 0.77 0.62
user 5 -0.04 -0.62 -0.70 -0.55 1.00 -0.24 0.22 -0.54 -0.54 -0.21 -0.22
user 6 0.53 0.56 0.28 0.52 -0.24 1.00 0.32 0.50 0.55 0.65 0.47
user 7 0.79 0.31 0.00 0.56 0.22 0.32 1.00 0.57 0.58 0.78 0.51
user 8 0.72 0.85 0.35 1.00 -0.54 0.50 0.57 1.00 1.00 0.77 0.62
user 9 0.76 0.84 0.34 0.99 -0.54 0.55 0.58 1.00 1.00 0.76 0.63
user 10 0.65 0.68 0.35 0.77 -0.21 0.65 0.78 0.77 0.76 1.00 0.62
Table 6.14: Correlation values of different products of Printer
User 1 user 2 user 3 user 4 user 5 user 6 user 7 user 8 user 9 user 10 Average
User 1 1.00 0.77 0.77 0.60 0.95 0.88 0.02 0.89 0.81 0.82 0.75
user 2 0.77 1.00 0.90 0.47 0.82 0.67 -0.48 0.76 0.88 0.87 0.67
user 3 0.77 0.90 1.00 0.61 0.83 0.70 -0.42 0.77 0.93 0.92 0.70
user 4 0.60 0.47 0.61 1.00 0.70 0.62 -0.29 0.56 0.47 0.45 0.52
user 5 0.95 0.82 0.83 0.70 1.00 0.93 -0.03 0.89 0.85 0.84 0.78
user 6 0.88 0.67 0.70 0.62 0.93 1.00 0.16 0.76 0.68 0.67 0.71
user 7 0.02 -0.48 -0.42 -0.29 -0.03 0.16 1.00 -0.09 -0.25 -0.24 -0.06
user 8 0.89 0.76 0.77 0.56 0.89 0.76 -0.09 1.00 0.87 0.83 0.72
user 9 0.81 0.88 0.93 0.47 0.85 0.68 -0.25 0.87 1.00 0.99 0.72
user 10 0.82 0.87 0.92 0.45 0.84 0.67 -0.24 0.83 0.99 1.00 0.72
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
149 | P a g e
Table 6.15: Correlation values of different products of Head Phone
User 1 user 2 user 3 user 4 user 5 user 6 user 7 user 8 user 9 user 10 Average
User 1 1.00 0.68 0.67 0.87 0.87 0.96 -0.87 0.81 0.85 0.83 0.67
user 2 0.68 1.00 0.90 0.84 0.84 0.67 -0.48 0.76 0.88 0.87 0.70
user 3 0.67 0.90 1.00 0.88 0.88 0.76 -0.88 0.75 0.85 0.92 0.67
user 4 0.87 0.84 0.88 1.00 1.00 0.90 -1.00 0.89 0.98 0.94 0.73
user 5 0.87 0.84 0.88 1.00 1.00 0.90 -1.00 0.89 0.98 0.94 0.73
user 6 0.96 0.67 0.76 0.90 0.90 1.00 -0.90 0.77 0.85 0.82 0.67
user 7 -0.87 -0.48 -0.88 -1.00 -1.00 -0.90 1.00 -0.89 -0.98 -0.94 -0.69
user 8 0.81 0.76 0.75 0.89 0.89 0.77 -0.89 1.00 0.89 0.87 0.67
user 9 0.85 0.88 0.85 0.98 0.98 0.85 -0.98 0.89 1.00 0.94 0.72
user 10 0.83 0.87 0.92 0.94 0.94 0.82 -0.94 0.87 0.94 1.00 0.72
Table 6.16: Correlation values of different products of Tablet
User 1 user 2 user 3 user 4 user 5 user 6 user 7 user 8 user 9 user 10 Average
User 1 1.00 -0.02 -0.03 0.03 0.23 0.22 0.04 0.12 0.00 -0.07 0.15
user 2 -0.02 1.00 0.73 0.38 0.81 0.61 -0.92 0.62 0.89 0.93 0.50
user 3 -0.03 0.73 1.00 0.31 0.76 0.70 -0.87 0.83 0.92 0.89 0.52
user 4 0.03 0.38 0.31 1.00 0.70 0.43 -0.54 0.47 0.39 0.43 0.36
user 5 0.23 0.81 0.76 0.70 1.00 0.64 -0.89 0.64 0.83 0.84 0.55
user 6 0.22 0.61 0.70 0.43 0.64 1.00 -0.75 0.71 0.79 0.65 0.50
user 7 0.04 -0.92 -0.87 -0.54 -0.89 -0.75 1.00 -0.79 -0.96 -0.98 -0.57
user 8 0.12 0.62 0.83 0.47 0.64 0.71 -0.79 1.00 0.87 0.77 0.52
user 9 0.00 0.89 0.92 0.39 0.83 0.79 -0.96 0.87 1.00 0.95 0.57
user 10 -0.07 0.93 0.89 0.43 0.84 0.65 -0.98 0.77 0.95 1.00 0.54
Table 6.17:Correlation values of different products of Smart Phone
User 1 user 2 user 3 user 4 user 5 user 6 user 7 user 8 user 9 user 10 Average
User 1 1.00 0.75 0.77 0.66 0.96 0.98 -0.01 0.98 0.87 0.84 0.78
user 2 0.75 1.00 0.73 0.75 0.73 0.72 -0.31 0.75 0.81 0.83 0.68
user 3 0.77 0.73 1.00 0.56 0.82 0.77 -0.42 0.83 0.90 0.92 0.69
user 4 0.66 0.75 0.56 1.00 0.72 0.64 -0.13 0.73 0.47 0.54 0.59
user 5 0.96 0.73 0.82 0.72 1.00 0.95 -0.02 0.99 0.83 0.85 0.78
user 6 0.98 0.72 0.77 0.64 0.95 1.00 0.01 0.96 0.84 0.81 0.77
user 7 -0.01 -0.31 -0.42 -0.13 -0.02 0.01 1.00 -0.03 -0.24 -0.24 -0.04
user 8 0.98 0.75 0.83 0.73 0.99 0.96 -0.03 1.00 0.84 0.84 0.79
user 9 0.87 0.81 0.90 0.47 0.83 0.84 -0.24 0.84 1.00 0.98 0.73
user 10 0.84 0.83 0.92 0.54 0.85 0.81 -0.24 0.84 0.98 1.00 0.74
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
150 | P a g e
Table 6.18:List of users which are excluded after user‟s sincerity analysis
Items User‟s list
Tablet User7
Laptop User5
Smartphone User7
Printer User7
Head Phone User7
6.6.4 Product Preference Score
User‟s sincerity measure removes the discrepancy in the data if available. Now, we
set a criterion of preference of the user. The value of ϕ obtained in equation 6.2 gives
the importance of a product in the eyes of a user. In [19] the author used a similar
formula and did not set any criteria of preference and consider all those products as
user‟s preferred products for which user just visited to look at the reviews, however, a
user can just visit the page for various reasons and not necessarily the visit indicates
that products is conforming the choice of the user. Therefore, we formulate the
criteria of preference as follows.
We consider following assumptions to set criteria of preference:
1) There may be the situation that the user does not need to save or email the
reviews or non-availability of printer may cause of no document being printed
even once. Thus the Boolean variables E, P and S which are used in equation
6.2, they all may be zero for review of a product which may be preferred by
the users.
2) If a person visits the link of the review presented before him for various
product in first six clicks i.e. the sequence of the visit to the link does not
exceed 6 out of the 10 products for each items which gives the 60% chances of
being visited. We say that the product has an importance in the eye of the user.
i.e. Vijmaygain value 1,2,3,4,5 or 6 only, it means -
(1/Vij) ≥ 0.16 ----------------- (A)
3) The time taken to read the review should be greater than 40 seconds, as we
consider that a review consists of 400 bytes on an average. Further we
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
151 | P a g e
speculate the reading speed of a user is 10 bytes/second. [22]. Refer to
equation (6.2), the value of t≥40 i.e. tij ≥2.
And Tij ≥ tij/tmax;
Tij ≥ 2/5;
Tij ≥ 0.4 ---------- (B)
Putting the values of in-equations A and B in equation (6.2), and considering
P=S=E=0; we get
ϕ ij≥ 0+0+0+0.16+0.4;
ϕ ij ≥0.56
From equation (6.3), we get:
δij ≥ 0.112 ---------- ( C )
Thus we define a function preferred (i,j) to set criteria of preference as;
1, if δ ij≥ 0.112
Preferred (i, j) =
0, otherwise
The criteria of preferences may be written as:
If preferred (i,j) = 1, product is preferred by the customers. If preferred (i,j) = 0,
product is not preferred by the customers. Thus, the product whose normalized
quantified vector score is greater than 0.112 will be considered as customer‟s
choice otherwise be neglected. We tabulate the criteria of preference in Table
6.19.
Table 6.19:Criteria of preference for a product to be preferred by a user
User‟s choice Value of preferred (i,j)
Preferred 1
Not preferred 0
In Table 6.20, the NPIS is shown for Laptop. User5 is excluded from the process
as the user was identified as insincere. The values of NPIS for various products which
are less than the threshold value, i.e. 0.112, are marked in bold.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
152 | P a g e
Table 6.20: Normalized Products Importance Score for Laptop
Average
Score Of
Products
User 1 User 2 User 3 User 4 User 6 User 7 User 8 User 9 User
10
L1 0.92 0.62 0.186 0.96 0.62 0.72 0.72 0.72 0.56
L2 0.53 0.68 0.52 0.62 0.76 0.46 0.62 0.66 0.66
L3 0.58 0.36 0.38 0.386 0.186 0.385 0.386 0.386 0.36
L4 0.346 0.53 0.53 0.33 0.57 0.348 0.37 0.37 0.386
L5 0.36 0.508 0.352 0.32 0.08 0.545 0.32 0.36 0.37
L6 0.302 0.512 0.52 0.312 0.305 0.302 0.312 0.112 0.312
L7 0.272 0.386 0.105 0.268 0.102 0.3 0.108 0.108 0.108
L8 0.308 0.305 0.348 0.265 0.1 0.312 0.105 0.105 0.105
L9 0 0.102 0.302 0.062 0.16 0.36 0 0 0.342
L10 0.509 0.1 0.3 0.06 0.352 0.37 0 0.102 0.34
6.6.5 User Personalized Ranking
Once we set criteria for preference, we will be getting a series of products that
customer prefer by ordering the value obtained by each user in descending order
where we will be having products for which the function preferred (i,j) =1. This may
lead to a partial list. The ranking of laptop based on product preference score (PPS)
i.e. the NPIS greater than the threshold value, for all the concerned users, is depicted
in Table 6.21. It is a partial list.
Table 6.21: Ranking of laptop by different users based on product preference score
Ranked
position User 1 User 2 User 3 User 4 User 6 User 7 User 8 User 9
User
10
1 L1 L2 L4 L1 L2 L1 L1 L1 L2
2 L3 L1 L2 L2 L1 L5 L2 L2 L1
3 L2 L4 L6 L3 L4 L2 L3 L3 L4
4 L10 L6 L3 L4 L10 L3 L4 L4 L5
5 L5 L5 L5 L5 L6 L10 L5 L5 L3
6 L4 L7 L8 L6 L3 L9 L6 L6 L9
7 L8 L3 L9 L7 L9 L4 L7 - L10
8 L6 L8 L10 L8 L5 L8 L8 - L6
9 L7 - L1 - - L6 - - -
10 - - - - - L7 - - -
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
153 | P a g e
Algorithm 6.1: Positional Rank aggregation
1: Repeat for x=1 to m
{
2: Repeat for y=1 to n
{
3: find ranked position of product ‟x‟ in the ranking of user ‟y‟, say it is ‟ k‟; k ∈ [0, n]
4: If (k! =0) /* ranking of product „x‟ is present for user „y‟*/
{
5: compute score „S(x,y)‟ for product „x‟ by user „y‟:
S(x,y) = [(m+k) – {(2*k)- 1}] ;
}
6: else /* product „x‟ is missing in the ranking of user „y‟ i.e. k=0 */
{
S(x,y) =0;
}
}
7: ( , )n
x
y i
S S x y
8: Sort product „x‟, x ∈ [1, n] in descending value of Sx; this arrangement will give ranked list of
products by user „y‟
}
We apply rank aggregation algorithm to get a single list that may be considered as the
final ranking by the user. We give the algorithm to find the aggregated ranking of the
products by different users. If we have „m‟ different products and products acquired
different positions in the ranking given by respective users. If the users involved in
the ranking process are „n‟. The procedure is represented in algorithm 6.1;
6.7 Results and Discussions
We have two different rankings for several products of 5 different items. First one is
the system ranking and another is the users‟ aggregated ranking. System ranking is the
final ranking recommended in [6] and the users‟ aggregated ranking is obtained by the
proposed comprehensive approach discussed in section 6.5. We evaluate the system
ranking on the basis of user‟s aggregated ranking. The values of the different
measures for comprehensive approach are discussed in section 6.7 and the discussion
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
154 | P a g e
of the results of the comparison of the proposed approach and other existing
techniques is performed in section 6.8.
We have adopted various measures to evaluate the system using proposed
comprehensive approach (C.A). The evaluation measures frequently used in to
evaluate recommender system are MRR, MAP, p@k, FPR@k, FNR@k, and
spearman rank correlation. These measures are employed and the recommender
system [6] is compared by the proposed approach. Thus, these measures help us in
evaluating the recommender system under evaluation, proposed in [6]. Each measure
is discussed separately in the subsequent sections. The values obtained for these
measures are tabulated and their pictorial representations are also shown in the
respective sub sections.
6.7.1 Mean Reciprocal Rank obtained using Comprehensive Approach
The Mean Reciprocal Rank (MRR) is discussed in section 6.3.6. We find MRR of all
items for their respective ranked first products by using comprehensive ranking for
each item. The values of Reciprocal Rank (RR) for all the items are 1except laptop;
which implies that all first ranked products of different items (except laptop) in
system ranking are also ranked first by comprehensive ranking. The mean reciprocal
rank (MRR) comes out to be 0.9. The values are shown in Table 6.22 and pictorially
depicted in Figure 6.11.
Table 6.22:Mean Reciprocal Rank of first ranked product of different items
Products RR
Laptop 0.5
Head Phone 1
Smart Phone 1
Tablet 1
Printer 1
MRR 0.9
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
155 | P a g e
Figure 6.11: Mean Reciprocal Rank of top rank-position for respective items using
Comprehensive Approach.
6.7.2 Precision@10 obtained using Comprehensive Approach
The values of P@1 to P@10 of user‟s ranking for respective products are given. The
variation in the precision value for each top position can easily be noticed with the help
of Table 6.23, as we can see that P@1 for laptop is 0, it is because the first ranked
product in the system ranking gets a second position in the user‟s ranking and not the
first position, and hence at first position we get precision as zero, whereas the value of
P@1 for headphone, smart phone, tablet and printer is 1.
Table 6.23:values of precision at k, for different items
P@1 P@2 P@3 P@4 P@5 P@6 P@7 P@8 P@9 P@10
Laptop 0 1 1 1 1 1 0.86 0.875 0.89 1
Head Phone 1 1 1 0.75 1 1 1 1 1 1
Smart Phone 1 0.5 1 1 1 1 1 0.875 0.89 1
Tablet 1 1 1 0.75 1 1 1 0.875 0.89 1
Printer 1 1 1 1 1 1 1 0.875 0.89 1
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
156 | P a g e
Figure 6.12: P@k for different items using Comprehensive Approach
We have depicted P@k in Figure 6.12, for all k=1 to k=10 graphically to elaborate
the precision value for top 10 positions. It is very clear that recommended system gives
100% precise recommendation for top 3 and top 5positions.
6.7.3 Mean Average Precision obtained using Comprehensive Approach
The value of the MAP is given in Table 6.24and a graphical representation is also
shown in Figure 6.13. The high values of MAP for respective products show the good
quality of the recommender system.
Table 6.24: Mean Average Precision for different products
Products MAP
Laptop 0.8625
Head Phone 0.975
Smart Phone 0.9265
Tablet 0.9515
Printer 0.9765
Figure 6.13: Mean Average Precision using Comprehensive Approach
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
157 | P a g e
6.7.4 FPR@10 obtained using Comprehensive Approach
The FPR@1 to FPR@10 is shown in Table 6.25, we can see that FPR@3 and FPR@5
is coming out to be 0, i.e. for top 3 and top 5 positions the recommendation has no
false positive error. The zero value of FPR@k, for k=3 and k=5 does not mean that
the system is free from error but it simply indicates that due to the change in ranking
position of the products, the value of FPR@k is coming out to be zero.Whereas for
other values of k, we get non-zero values of FPR@k, it clarifies that system is not
biased as it exhibits error for other values of k. Also, the zero value of False Positive
Rate for different ranking positions represents the degree of preciseness of the system.
We define Average of FPR@k as follows;
1
@. FPR@k = ---------------------------------- (6.11)
k
i
FPR iAvg
k
The avg. FPR@k gives the more accurate measure of fallacy of the system. Avg.
FPR@5 and Avg. FPR@10 for the system is obtained and presented in Table 6.26.
Table 6.25:values of FPR@10 for different products
Laptop Head Phone Smart Phone Tablet Printer
FPR@1 1 0 0 0 0
FPR@2 0 0 0.5 0 0
FPR@3 0 0 0 0 0
FPR@4 0 0.25 0 0.25 0
FPR@5 0 0 0 0 0
FPR@6 0 0 0 0 0
FPR@7 0.14 0 0 0 0
FPR@8 0.125 0 0.125 0.125 0.125
FPR@9 0.11 0 0.11 0.11 0.11
FPR@10 0 0 0 0 0
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
158 | P a g e
Table 6.26: Avg. FPR@5 and Avg. FPR@10 for all the items
Laptop Head Phone
Smart
Phone Tablet Printer
Mean of
Average FPR
for all products
Avg
FPR@5 0.2 0.05 0.1 0.05 0 0.08
Avg
FPR@10 0.1375 0.025 0.0735 0.0485 0.0235 0.0616
Figure 6.14: Average FPR@5 using Comprehensive Approach
Figure 6.15:Average FPR@10 using Comprehensive Approach
The above value of the measure of false positive and false negative is depicted in
Figure 6.14andFigure 6.15. The values indicate the performance of the system which
is being evaluated using comprehensive approach.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
159 | P a g e
6.7.5 FNR@10 obtained using Comprehensive Approach
In Table 6.27, values of FNR for different positions are shown. For the corresponding
table which illustrates that FNR@3 and FNR@5 are zero; it shows the zero false
negative error for the recommender system for top 3 and top 5 positions respectively.
We define Average of FNR@k as follows;
1
@. FNR@k = ----------------------------------------- (6.12)
k
i
FNR iAvg
k
The FNR@k works similar to FPR@k in assessing the systems accuracy in terms of
its prediction to user‟s choices. These values of avg. FNR@k for k=5 and k=10 are
shown in Figure 6.16 andFigure 6.17 respectively.
Table 6.27: Values of FNR@10 for different products
Laptop Head Phone Smart Phone Tablet Printer
FNR@1 1 0 0 0 0
FNR@2 0 0 0.5 0 0
FNR@3 0 0 0 0 0
FNR@4 0 0.25 0 0.25 0
FNR@5 0 0 0 0 0
FNR@6 0 0 0 0 0
FNR@7 0.14 0 0 0 0
FNR@8 0.125 0 0.125 0.125 0.125
FNR@9 0.11 0 0.11 0.11 0.11
FNR@10 0 0 0 0 0
Table 6.28: Avg. FNR@5 and Avg. FNR@10 for all the items
Laptop Head Phone
Smart
Phone Tablet Printer
Mean of
Average FPR
for all products
Avg
FNR@5 0.2 0.05 0.1 0.05 0 0.08
Avg
FNR@10 0.1375 0.025 0.0735 0.0485 0.0235 0.0616
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
160 | P a g e
Figure 6.16: Average FNR@5 using Comprehensive Approach
Figure 6.17: Average FNR@10 using Comprehensive Approach
6.7.6 Spearman Correlation value using Comprehensive Approach
We find the spearman correlation coefficient between system ranking and user raking.
The values are shown in Table 6.29, anddepicted in Figure 6.18. It is evident from the
obtained values that both the ranking is highly correlated.
Table 6.29:Spearman correlation coefficient for different products
Products Spearman Correlation Coefficient
value
Laptop 0.9030
HeadPhone 0.9878
SmartPhone 0.9515
Tablet 0.9515
Printer 0.9636
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
161 | P a g e
Figure 6.18: Spearman correlation coefficient between system ranking and Comprehensive
Approach based ranking
6.8 Relative Performance of the Recommender Systems using Proposed
Comprehensive Approach and other Existing Evaluation Approaches
In this section we have discussed the details of the results obtained while comparing
the proposed comprehensive approach with existing Average Scoring based technique
[40] and Rank Aggregation based technique [39]. The section 6.7 gives the values of
calculated metric using comprehensive approach. The relative performance of the
system under evaluation, by using proposed comprehensive approach and other
related work [39], [40]is analyzed and their corresponding values are tabulated below.
In Table 6.30, those metric which show accuracy are indicated. The higher values
of these evaluation metrics imply the examined system performance is better. In Table
6.31, the evaluation parameters which measure errors are presented. The lower values
of these parameters indicate the least error occurring in the examined system.
We have obtained a Comprehensive Veracity Measure (CVM) for the system using
all three different approaches which help in assessing the performance of the system
under different scenario. CVM is given by;
Sum of values of different evaluation metrics ----------------- (6.13)
Total number of metrics
@5 Spearman correlation coefficient --- (6.14)
4
CVM
MRR P MAPCVM
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
162 | P a g e
Thus, CVM for Comprehensive Approach (C. A)isreferred as „CVM (CA)‟ and
calculated as;
CVM (CA) = (.9+1+.9384+.9514)/4
CVM (CA) = 0.94745
Table 6.30:Mean Reciprocal Rank, P@5, Mean Average Precision and Spearman correlation
coefficient for different approaches
Approach
Mean
Reciprocal
Rank (MRR)
P@5
Mean
Average
Precision
(MAP)
Spearman
Correlation
Coefficient
Comprehensive
Veracity measure
(CVM)
Proposed
Comprehensive
Approach
.9 1 .9384 0.9514 0.94745
Average Scoring
based technique
[40]
1 0.96 0.9501 0.9030 0.927275
Rank Aggregation
based technique
[39]
1 0.856 0.866 0.9363 0.940575
Figure 6.19: Comprehensive Veracity Measure of different approaches
Similarly, we find the CVM for previous approaches. These values of CVM for
average based scoring technique; rank aggregation approach and comprehensive
approach are shown in Table 6.30 and pictorially represented in Figure 6.19.
The results of MRR for the various evaluation measures indicate that the system
has recommended all the first ranked products of different items exactly as most of
the evaluation measures suggest. However, if we consider comprehensive approach,
we get MRR as 0.9, i.e. 90% of first ranked products are similar in system ranking
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
163 | P a g e
and comprehensive ranking. Average P@5 for all the items for different evaluation
approaches is also shown in Table 6.30. The maximum value of P@5 for the
recommender system has obtained by comprehensive approach. However, minimum
value of P@5 is .86, which is good enough for the recommender system to be
considered as accurate in terms of precision.
The Mean Average Precision (MAP) of all ranking positions for different
evaluation approaches has shown. Interestingly, the system has a very good MAP by
all the evaluation approaches. But the maximum MAP is obtained by Average
Scoring based Approach.
We have shown spearman correlation of all the evaluation approaches with the
recommender system under study. It is evident from the value that all the approaches
are highly correlated; however, comprehensive approach is more correlated than the
other approaches. The high correlation of system ranking with the ranking of other
approaches clearly indicates that the system ranking is very close to the user‟s choice
and good enough to be chosen as a suggested recommender system for the
recommendation of various products.
Also, observing the „comprehensive veracity measure‟, it is found that the all 4
metrics measuring veracity of the system has approximately near values. The
proposed approach with „rank aggregation based approach‟ shows very similar results
however the average scoring based technique differs slightly. The reason behind is
obvious, as both the methods incorporate aggregation algorithm of the ranked items of
the users, whereas the average scoring technique relies upon numerical value assigned
to different products and hence differ significantly. Since the small data set is used in
the work, on a larger data set the comprehensive approach would evaluate the system
veracity more accurately as the probability of users being insincere can be determined
and other two methods would not be able to trace these issues.
Similarly, the measure of fallacy, FPR@k and FNR@k gives how much the system
can defense the user‟s dissatisfaction and avoid the user-reluctant tendency in the
recommendation.
In Table 6.31, the measures of FPR@5 and FNR@5 for the different approaches
have been listed. Since the preference criteria and user‟s sincerity is not included in
the two studies [40], [39], the evaluation approach is error prone and identify the error
which basically is due to the outlier existence in the data. Therefore, average scoring
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
164 | P a g e
based technique encounters maximum FPR@5 which is 0.144; interestingly the
veracity measure for the technique is also high, i.e. a contradiction. The measures
highlight the loop hole in the previous approach which is overcome by the proposed
comprehensive approach. Thus it can be clearly concluded that proposed
Comprehensive approach outperforms the other existing techniques.
Also, FPR@5 is not the overall error score but it just indicates for top k positions
how precise the recommendation is. Thus, the value of FPR@k for different k
represents the degree of biasness that the system has. If FPR@k remains same for
each k, it implies that the system is biased and does not perform with consistency. The
values are shown in Figure 6.20andFigure 6.21.
Table 6.31:FPR@5 and FNR@5 for different approaches
Approach Average FPR@5 for all items Average FNR@5 for all items
Proposed Comprehensive
Approach 0.08 0.08
Average Scoring based
technique[40] 0.144 0.144
Rank Aggregation based
technique[39] 0.04 0.04
Figure 6.20: Average FPR@5 for all items
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
165 | P a g e
Figure 6.21: Average FNR@5 for all items
6.8.1 Comparison of Proposed Comprehensive Approach with Existing
Evaluation Strategies
The present work suggests an approach to evaluate the recommender system. The
evaluation is based on implicit feedback. The local dataset is used which is created by
observing the users‟ behavior and the implicit record of their activity over the
provided link and reviews of the items. The users behavior is noticed for those
selected items which have been sent to them and taken from the dataset described by
the author [243].
The different evaluation studies have been reported in the literature and the
respective strategies are discussed in section 6.2. We do not have any mathematical
model or simulation technique to compare the evaluation approaches with each other
having so diverse data and extensive approaches. Thus we have chosen different
factors which have been incorporated in the proposed evaluation system. These
factors have been used by others evaluation approaches as well. The advantages of
employing these techniques and why the proposed approach should be preferred is
discussed below. In the Table 6.32, a comparative study is presented which illustrates
how proposed approach has advantages over the evaluation techniques existing in the
literature. Seven different factors have been considered. From the table it is evident
that all the existing evaluation studies have proposed a new framework for evaluation
or have been suggested a new metric to evaluate the system. Only Herlocker et al.
[186] has not proposed new framework, however, they have studied the evaluation
scheme for RS thoroughly and have provided a detail discussion for these approaches
including more than 6 metrics for the purpose.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
166 | P a g e
Table 6.32: Comparison of proposed Comprehensive Approach with existing evaluation
strategies
Evaluation
approaches
New
metric/fram
ework
proposed
No. of
metric
s used
≥6
Sincerity
check of
users
Mathematical
formulation
for criteria of
preference
Explicit/I
mplicit
feedback
from
users
Experimental
analysis with
existing RS
Schroder et al.
[266]
Herlocker et al.
[186]
Sohail et al.
[39], [40]
Olmo&Gaudios
o[267]
Cremonesi et al.
[268]
Shani and
Gunawardana[5
1]
Proposed
Comprehensive
Approach
Almost all the metrics have used six or more than six metrics in the evaluation
process. However, Sohail et al. [40] have used less than 6 metrics. Also, Olmo &
Gaudioso[267] have used less than six metrics. We have proposed the system and
have also used 6 metrics to evaluate the system. Five from seven approaches have
also employed explicit/implicit feedbacks from users. Either the techniques have
described in details about how they have incorporated feedbacks or they have
themselves collected feedbacks from users.
In the same way, not all the techniques have experimentally evaluated any system
by their proposed approach. In the previous work, authors have evaluated the system
but preferably not discussed the differences in their adopted approach with the
approaches have been used by researchers in the literature. The point where the
proposed approach takes advantages over other is the sincerity check of the users and
well-defined criteria for defining the preference criteria of the users. The proposed
Comprehensive Approach (CA) includes all the factors whereas no other evaluation
strategy provides user sincerity check and only one has the provision which explicitly
defines the criteria of the selection and advices the mechanism to decide threshold.
This clearly designates the superiority of the proposed approach.
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
167 | P a g e
6.9 Summary
In this chapter, we have put forward two different user feedback based framework for
the evaluation of Recommender Systems (RS). The first framework depends upon
explicit feedback whereas another one utilizes implicit feedback. The evaluation of
RS based on explicit feedback is used to evaluate the book recommendation
techniques employed in this work and discussed in chapters 3, 4 and 5. The implicit
feedback based evaluation mechanism is used to evaluate RS presented in [243]. The
reason behind choosing explicit feedback for the evaluation of book recommender
systems proposed in this work is integrity of the experts whose feedbacks are taken
and considered as a base in the evaluation process, whereas the system proposed in
[243] is meant for general purpose of daily needs commodities and hence user‟s
sincerity and authenticity must be examined. This is why implicit feedback
mechanism is applied over it.
Through implicit feedback, a comprehensive approach for the evaluation of
recommender systems is suggested. The proposed methodology tries to measure
sincerity of the users who provide their feedback for the evaluation of the
recommender system. The computation of user‟s sincerity measure eliminates the
feedback of insincere users and in turn, makes the evaluation process reliable. Further,
the methodology outlines a procedure to decide whether a specific product, whose
review site is visited by the user, is to be considered as a product preferred by the user
or not. Hence, the proposed evaluation approach is poised to be a fairly realistic
approach and better than the other related evaluation techniques, which do not have
any provision to measure user‟s sincerity and which consider any product, whose
review site is visited by the user, as a product preferred by the user.
The proposed comprehensive approach is used to evaluate the performance of a
recommender system and the result of the evaluation is presented. We compare the
aggregated ranking of products obtained in the comprehensive evaluation approach
with the recommender system‟s ranking of the products, and compute the values of a
good number of evaluation metrics, namely Spearman correlation coefficient,
precision at k, Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), false
positive rate (FPR) and false negative rate (FNR). We also evaluate the performance
of the recommender system using two other related evaluation techniques proposed in
[19, 18] and compute the values of the same set of evaluation metrics. Since we do
Chapter 6: Evaluation of Recommender Systems
___________________________________________________________________________________
168 | P a g e
not have the true ranking of the products recommended by the recommender system,
we cannot decide objectively which of the three evaluation techniques is able to
evaluate the recommender system most accurately. Hence, the values of the
evaluation metrics obtained using the three evaluation techniques are compared and
the results of this comparison is pictorially represented.
Since the comprehensive evaluation approach is a fairly realistic approach as
discussed above, the high values of Spearman correlation coefficient, precision at k,
Mean Average Precision (MAP), Mean Reciprocal Rank (MRR) and low values of
false positive rate and false negative rate, obtained for the comprehensive evaluation
approach, clearly indicates that the recommender system under evaluation performs
well. Hence, the products recommended by the recommender system may satisfy the
user and the user may purchase them.
As more than 75 experts are contacted for 100 books. Few of them were given
more than 1 course. These people are in several universities in India, KSA, Iraq, Iran,
Jordan, and USA. Also few of them are in leading tech companies in above countries.
That‟s why we have selected explicit feedback evaluation scheme for the examining
the proposed book recommendation techniques.
The final result for the methodology which utilizes explicit feedback is
summarized in the respective tables. Table 6.16 shows the values of parameters which
have estimation of errors, and table 6.17 has the values of parameters indicating
precision. The technique which has higher value of these parameters would be treated
as better one.
The results show that amongst the different approaches used for the
recommendation of books, namely PAS, OWA with quantifiers „most‟, „at least half‟
and „as many as possible‟, ORWA and opinion mining; the most preferred approach
by the evaluation of experts is book recommendation using opinion mining technique.
169 | P a g e
Chapter 7
Conclusion and Future Direction
7.1 Introduction
The research work carried out in the thesis aimed at exploring the role of opinion
mining, a sub branch of web mining, in recommender systems and how it can
overcome the prevailing issues in the concerned techniques. In this last chapter,
research findings and our contributions are summarized; also future direction for the
research is highlighted. In section 7.2, the concluding remarks of the different works
carried out in respective chapters are discussed. The section describes the pros and
cons of the adopted approaches. The variation in the results of recommendation due to
the change in the technique is addressed. In section 7.3, the scope of the future work
and limitation of the adopted techniques are suggested.
7.2 Conclusion
We have reviewed the state of the art in recommender systems. A detail description of
the various techniques with the diagrammatic representations and common examples,
are given in Chapter 2. These details are easy to understand the approaches adopted
in recommender systems design. The contributions of the researchers on the topic are
focused and their relative comparison is discussed. The study also reveals the major
flaws with the leading existing techniques.
Suggestion to overcome the drawbacks which were identified at the time of the
study of literature is proposed with an intention of introducing opinion mining as a
solution. Not only the method but we have also suggested a comprehensive approach
for evaluation of the recommender system. In Chapter 3, we have introduced a rank
aggregation algorithm based recommendation of books; we call it „Positional
Aggregation based Scoring (PAS) technique‟. The link mining concepts are
incorporated to find the top raked universities recommendations for the books under
their prescribed syllabus and the aggregation scheme is employed to aggregate these
ranking and recommend top books for students.
Chapter 7: Conclusion
__________________________________________________________________________________
170 | P a g e
A fuzzy based aggregation operator, OWA (Ordered Weighted aggregation), have
been utilized and discussed in Chapter 4. OWA is implemented to give a variety of
experiments and verifications of proposed techniques. In order to rely more upon
voters and rankers prestige, an „Ordered Ranked Weighted Aggregation (ORWA)‟ is
suggested for the book recommendation.
The Ordered Ranked Weighted Aggregation incorporates rank of the rankers to
emphasize the importance of the rankers as a book recommended by best ranked
institution must get high preference than a book which is recommended by a lower
ranked institution. The ORWA gives the ranking positions of the recommended
books, along with the total recommended books. The strength of assigning weights to
the rankers in the ORWA provides a better recommendation. We believe the
proposed technique may meet the user‟s need and provide them the perfect books they
need.
An extensive approach based on opinion mining is also proposed. The OWA and
ORWA used recommendations from universities authorities, i.e. experts. It seems
adequate to involve the users rather than only the experts, for a better understanding
of their preferences and what they actually love to have? To observe the user‟s
requirements and their like or dislike opinion about an item, particularly books,
opinion mining techniques have been applied. The recommendation of books based
on opinion mining is presented before the users. All these suggested approaches are
evaluated to show the best amongst all.
We have put forward two different user feedback based frameworks for the
evaluation of Recommender Systems (RS). The first framework depends upon
explicit feedback whereas another one utilizes implicit feedback. The evaluation of
RS based on explicit feedback is used to evaluate the book recommendation
techniques employed in this work and discussed in chapters 3, 4 and 5. The implicit
feedback based evaluation mechanism is used to evaluate RS presented in [243]. The
reason behind choosing explicit feedback for the evaluation of book recommender
systems proposed in this work, is the integrity of the experts whose feedbacks are
taken and considered as a base in the evaluation process, whereas the system proposed
in [243] is meant for general purpose of daily needs commodities and hence user‟s
sincerity and authenticity must be examined. This is why implicit feedback
mechanism is applied over there.
Chapter 7: Conclusion
__________________________________________________________________________________
171 | P a g e
The comprehensive approach to evaluate recommender systems which is based on
user feedback tries to measure sincerity of the users who provide their feedback for
the evaluation of the recommender system. The computation of user‟s sincerity
measure eliminates the feedback of insincere users and in turn, makes the evaluation
process reliable. Further, the methodology outlines a procedure to decide whether a
specific product, whose review site is visited by the user, is to be considered as a
product preferred by the user or not. Hence, the proposed evaluation approach is
poised to be a fairly realistic approach and better than the other related evaluation
techniques, which do not have any provision to measure user‟s sincerity and which
consider any product, whose review site is visited by the user, as a product preferred
by the user.
The proposed comprehensive approach is used to evaluate the performance of a
recommender system and the result of the evaluation is presented. We compare the
aggregated ranking of products obtained in the comprehensive evaluation approach
with the recommender system‟s ranking of the products, and compute the values of a
good number of evaluation metrics, namely Spearman correlation coefficient,
precision at k, Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), false
positive rate (FPR) and false negative rate (FNR). We also evaluate the performance
of the recommender system using two other related evaluation techniques proposed in
[19, 18] and compute the values of the same set of evaluation metrics. Since we do
not have the true ranking of the products recommended by the recommender system,
we cannot decide objectively which of the three evaluation techniques is able to
evaluate the recommender system most accurately. Hence, the values of the
evaluation metrics obtained using the three evaluation techniques are compared and
the results of this comparison is pictorially represented. The said method can be
served as a base to evaluate the recommender systems performance based on explicit
feedbacks.
The explicit feedback based evaluation of the present study is performed. The
results of all the proposed schemes, i.e. PAS, OWA with quantifiers „at least half‟, „at
most‟ and „as many as possible‟, ORWA and Opinion mining are shown and
compared. The comparison suggests the advantage of using opinion mining
techniques over other approaches. Also, user feedback is the basis of the evaluation of
the proposed scheme; hence the work also gives a direction of utilizing feedback from
Chapter 7: Conclusion
__________________________________________________________________________________
172 | P a g e
users for evaluation process. There is a great deal of future enhancement in the work,
which is discussed in the next section.
7.2 Future Directions
The work presented in this thesis has comprehensively covered web mining and soft
computing techniques for the recommendation of books and evaluation of
recommender systems, both. Still, there is a lot of scope to enhance the work and
there are several areas for which the present works gives the direction to explore.
These future directions are listed below.
i. The different approaches discussed in this thesis are specific to the selected
domain of books and products. In future, the work can easily be extended for
the enlarged dataset.
ii. The present work is designed in Indian perspectives. Hence, the syllabus of the
books is taken from top universities of India only. In future, the approach can
be comfortably implemented to any institute and any country by considering
the universities around the world.
iii. The proposed opinion mining techniques exploits features based
recommendations. We have selected different features of books depending
upon users‟ interests. For simplicity, relatively less number of features is
selected. In future, the features selection procedure can be modified to
increase or decrease the total number of features according to the interests of
the users and item types, both.
iv. Our procedure for checking of the users‟ sincerity is solely based upon the
assumption that majority of the users are sincere. Also, there could be no
genuine users or less sincere users. In such a case much cannot be done by our
method. In future, the sincerity checking for the users in similar situations can
be formulated and applied to deal with.
Chapter 7: Conclusion
__________________________________________________________________________________
173 | P a g e
v. Our focus in this work is to provide a framework for the recommendation of
online items, especially books. Thus instead of emphasizing on the spam
detection in reviews, we just concentrated on how these reviews can be
formulated to make appropriate recommendations. Hence, the spam detection
is not well studied. Therefore, in future it would be interesting to see what if
the customers‟ reviews can be checked for spam and only the spam filtered
reviews are involved in the recommendation process.
vi. Further, the user emotion state can be applied to weight the feature according
to the emotional condition of the users. The emotions can be observed by
capturing facial expressions, prior events happening with the users, etc. this
emotional state would help in analyzing the review genuineness and how
much it can be relied upon?