chapter 3 analysis of existing system and...
TRANSCRIPT
3. Analysis of Existing System and Limitation 43
CHAPTER 3
ANALYSIS OF EXISTING SYSTEM AND LIMITATION
3.1 REVIEW PREVIOUS RESEARCH FINDINGS AND COMPARATIVE STUDY
1. Fab: Content-based, collaborative recommendation
Author: M. Balabanovi´c and Y. Shoham
Year: 1997
Approach: Content-boosted
Summary:
In this paper, authors have combined CB and CF methods. This designing solves Scalability
Problem. Both CB and CF have several limitations. Pure content-based approach has several
disadvantages like Over-specialization; only capture certain aspects/features of product. While
pure CB has also its own disadvantages like new item/user problem, data sparsity.
Fab System[30]: The process of recommendation can be partitioned into two steps:
1. Collection of items to form a manageable database or index, and
2. Subsequently selection of items from this database for particular users.
FAB has three main Components: 1.Selection Agent- Search pages for specific users 2.
Collection Agent –Search pages for specific topic 3. Central route Role of Agents in FAB:
Collection agents send pages, which are founded during process, to central router. Central
Router sends them to those users whose profiles match with it. Selection Agent discards the
pages which are already seen by users.
When the user has requested, received, and looked over their recommendations, they are
required to assign appropriate ratings from a 7-point scale. This rating is stored in user’s
personal agent profile for further recommendation and it is also forwarded to Collection agent’s
profile. The collection agents’ profiles represent a topic of interest to a dynamically changing
group of users, as opposed to a user’s profile, which represents multiple interests possibly
served by several collection agents. In future work, mainly two research issues are to be
tackled. They aim to study the effects of massively scaling up the number of users, and they
plan to continue the investigation of the dynamic processes involved.
3. Analysis of Existing System and Limitation 44
2. Recommendation as classification: using social and content-based Information in
recommendation
Author: Chumki Basu, Haym Hirsh, and William Cohen
Approach: CF and content-based information
Year: 1998
Summary:
In this paper, author have explained an inductive learning approach to recommendation that
is able to use both ratings information and other types of information about each product
in predicting user preferences. Here, author has used hybrid features[25] that combine
elements of social and content-based information makes it possible to achieve more
accurate recommendations.
3. Combining content-based and collaborative filters in an online newspaper
Author: Claypool, M., Gokhale, A., Miranda, T., Murnikov, P., Netes, D., & Sartin
Year: 1999
Application: Online Newspaper
Summary
Authors have presented new approach that combines content filters with the depth of
collaborative filtering. Pure collaborative filtering suffers from early rater problem, sparsity
problem and gray Sheep problem. CB doesn’t suffer from above problem but it has its own
disadvantage like it has difficulty to distinguish between high and low quality content. Further,
number of items in content category increases, it decreases the effectiveness of CB approach.
Here it combines collaborative filtering prediction with content based prediction using a
weighted average. CF gives inaccurate result in case of lack of history data and CB gives
inaccurate result in case where users have not specified explicit keywords. It is proven in
Experimental study that combination of CB and CF give more accuracy(minimum MAE)[26]
in comparison of pure CB and pure CF approach.
In future work, this approach can be extended by adding demographic approach to improve
accuracy and to work more on prediction strength of CF.
3. Analysis of Existing System and Limitation 45
4. Combining collaborative filtering with personal agents for better recommendation
Author: Nathaniel Good, J. Ben Schafer, Joseph A. Konstan, Al Borchers,
Badrul Sarwar, Jon Herlocker, and John Riedl
Year: 1999
Summary
Authors have presented in this paper that a CF framework can be used to combine personal IF
agents and the opinions of a community of users to produce better recommendations than either
agents or users can produce alone.
Four key models are presented here [20-jour]:
1. Pure collaborative filtering using the opinions of other community members
2. A single personalized "agent" – a machine learning or syntactic filter
3. A combination of many "agents"
4. A combination of multiple agents and community member opinions
Four primary hypotheses are given below:
1. H1: The opinions of a community of users provide better recommendations than a
single personalized agent.
2. H2: A personalized combination of several agents provides better recommendations
than a single personalized agent.
3. H3: The opinions from a community of users provide better recommendations than a
personalized combination of several agents.
4. H4 : A personalized combination of several agents and community opinions provides
better recommendations than either agents or user opinions alone
The most important results they have found were the value of combining agents with CF and of
combining agents and users with CF. In essence, these results suggest that an effective
mechanism for producing high-quality recommendations is to throw in any available data and
allow the CF engine to sort out which information is useful to each user. In effect, it becomes
less important to invent a brilliant agent; instead we can simply invent a collection of useful
ones. To take advantage of learning agents, these engines must be redesigned to accommodate
"users" with dynamic rating habits.
They have examined several different CF engine designs that could efficiently use filterbots.CF
outperformed linear regression as a combining mechanism for agents. While linear regression
3. Analysis of Existing System and Limitation 46
should provide an optimal linear fit, it appears that CF's non-optimal mechanism actually does
a better job avoiding over fitting the data when the number of columns approaches the number
of rows.
CF also has the advantage of functioning on incomplete (and indeed very sparse) data sets,
suggesting that it retains its value as a useful combination tool whenever human or agents are
unlikely to rate each item.
Future work should both incorporate larger user sets (other experiments have consistently
shown MAE values in the range of 0.71-0.73 and ROC sensitivity values near 0.72 for movie
lens communities with thousands of users) and look explicitly at closer-knit communities to see
whether a smaller but more homogeneous community would have greater benefits from
collaborative filtering.
In the future, they plan to examine further combinations of users and agents in recommender
systems. In particular, they are interested in developing a combined community where large
numbers of users and agents co-exist. One question they hope to answer is whether users who
agree with each other would also benefit from the opinions of each other's trained agents.
5. PTV: Intelligent Personalized TV Guides
Title: PTV: Intelligent Personalised TV Guides
Author: P. Melville, R.J. Mooney, and R. Nagarajan
Year: 2002
Summary
PTV represents a convergence of technologies that provides an effective solution to the
very real problem of providing people with relevant TV listings information as digital TV
becomes a reality. PTV personalises TV information to meet the viewing preferences of
individual users by integrating two different information-filtering strategies, case-based
reasoning and collaborative filtering, with user profiling techniques.
The resulting hybrid personalisation technique allows programme recommendations to be
made according to the type of programmes a target user has enjoyed in the past as well as
the programmes that other similar users have enjoyed. In the near future, a WAP version of
PTV will be formally launched and they have believed a similar success story will unfold
as mobile phone users recognise the real benefits of high quality content personalisation on
3. Analysis of Existing System and Limitation 47
their restricted mobile handsets [27]. In fact, we argue that traditional TV listings services
are not appropriate given the screen and bandwidth limitations of the current generation of
WAP devices – a personalised service such as PTV is the best available solution. The PTV
systems are built around a content personalisation engine that can be readily adapted to
practically any source of information content and Changing Worlds is currently using this
technology to develop the next generation of intelligent, personalised information services.
6. Probabilistic models for unified collaborative and contentbased recommendation in
Sparse-data environments
Author: Alexandrin Popescu, Lyle H. Ungar, David M. Pennock, Steve Lawrence
Year: 2001
Summary
Authors have proposed a unified probabilistic framework for merging collaborative and
content-based recommendations.
This model incorporates three-way co-occurrence data by presuming that users are
interested in a set of latent topics which in turn “generate” both items and item content
information. Model parameters are learned using expectation maximization (EM), so the
relative contributions of collaborative and content-based data are determined in a sound
statistical manner.
Here presented three probabilistic mixture models for recommending items based on
collaborative and content-based evidence merged in a unified manner. Incorporating
content into a collaborative filtering system can increase the flexibility and quality of the
recommender. Moreover, when data is extremely sparse—as is typical in many real world
applications additional content information seems almost necessary to fit global
probabilistic models at all.
They have found that a particularly good way to include content information in the context
of a document recommendation system is to treat users as reading words of the document,
rather than the document itself. In the case, this increased the density from 0.38% to almost
9%[12], resulting in recommendations superior to ANN.
There are many areas for future research. Similar methods to those presented here might be
used to recommend items such as movies which have attributes other than text. A movie
3. Analysis of Existing System and Limitation 48
can be viewed as consisting of the director and the actors in it, just as a document contains
words. Both of the sparsity reduction techniques, similarity-based smoothing and an
equivalent of a user-words aspect model, can be used.
Em is guaranteed to reach only a local maximum of the training data log-likelihood.
Multiple restarts need to be performed if one desires a higher quality model. They have
planned to investigate ways to intelligently seed em to reduce the need for multiple restarts,
which can be costly when fitting datasets of non-trivial size.
The user-words model does not explicitly use the popularity of items. Including such
information may further improve the quality of the recommendations made by the model,
but requires additional work on combining and calibrating model predictions with
document popularity.
7. Content-boosted collaborative filtering for improved recommendations
Title: Content-Boosted Collaborative Filtering for Improved Recommendations
Author: P. Melville, R.J. Mooney, and R. Nagarajan
Year: 2002
Summary
Authors have merged Content-Collaborative filtering in a hybrid manner. This approach
overcomes disadvantages of CF systems by exploiting content information of the items
already rated. It mainly predicts recommendation even on sparse data. So basically it solves
cold start problem. Further, if rating history for several products is not available then also it
is not an issue here because it efficiently deals with sparse data.
CBCF Approach:
1. First Content predictor takes input as user rating (Sparse rating data).
2. CP convert sparse rating matrix into full rating matrix(pseudo user-ratings vector)
3. Pseudo user-ratings vector provides as input to CF.
4. CF gives top recommendation.
Experimental study proves that CBCF performs better than Pure CB and pure CF
algorithm, it also solves problem of first rater problem. It tackles data sparsity. Future work,
CBCF performs consistently better than pure CF, the difference in performance is not very
large (4%)[2].
3. Analysis of Existing System and Limitation 49
8. Clustering Approach for Hybrid Recommender System
Author: Qing Li, Byeong Man Kim
Year: 2003
Summary:
Authors have presented Clustering techniques to solve cold start problem. Mainly cold start
problem occurs when products are recommended using CF. CF cannot recommend new items
to users without any past rating and completely denies any information that can be extracted
from content of items. So here authors have proposed hybrid system which combines CB and
CF together.
Ichm – item based clustering hybrid model [28]:
1. First apply clustering algorithm to group the items, and then use the result, which is
represented by the fuzzy set, to create a group-rating matrix.
2. Compute the similarity: firstly, calculate the sub-similarity of group-rating matrix, than
calculate the sub-similarity of item-rating matrix. At last, the total similarity is the linear
combination of the above two.
3. Make a prediction for an item by performing a weighted average of deviations from the
neighbour’s mean.
After creating new rating matrix by grouping the items, next step is calculate the similarity. To
calculate similarity, firstly, it to use the Pearson co-relation based algorithm on item-rating
matrix then adjusted cosine algorithm also calculate similarity from the group-rating matrix. At
the end, total user similarity is a linear combination of the above two. Formulas are listed
below:
Pearson correlation-based similarity.
S(a,u) = ∑ ( – ̅̅ ̅̅ )( – ̅̅ ̅̅ )
√∑ ( – ̅̅ ̅̅ ) ∑ ( – ̅̅ ̅̅ )
Adjusted cosine similarity.
S(a,u)=∑
√∑
√∑
In last step Collaborative prediction use to predict top N product. The general formula for a
prediction on the item i of user k is:
3. Analysis of Existing System and Limitation 50
= ̅̅̅̅ +
∑ ( )
∑ | |
The items, which are falls in category of no history data, we can make predictions for users on
this item, based on the group-rating matrix.
9. Cinemascreen Recommender Agent:Combining Collaborative and Content-Based
Filtering
Author: James Salter and Nick Antonopoulos
Year: 2006
Summary:
Cra system:
Collaborative filtering first:
it involves first finding a subset of users with film tastes similar to the current user. Comparing
the current user’s rating history with the history of every other user, the system finds the
current user’s potential peers—that is, other users who have rated films the current user has
rated(pearson’s product-moment correlation coefficient, r).
Although this approach might generate a larger set of films for making recommendations, it
would likely also reduce the prediction accuracy. To make its predictions, our collaborative
filtering process uses the peer ratings and gives a weighted average to each film according to
the strength of each peer’s correlation with the current user.
Once all calculations are complete, the agent stores the list of films and predicted ratings. The
system also stores the number of significant peers who rated the film because it gives an
indication of the potential recommendation’s strength. The system can therefore use this
number as a secondary sorting field when it produces recommendation lists. The system then
feeds the predicted ratings into the content-based filtering algorithms.
Content-based filtering on collaborative results [29]
We designed the content-based filtering to use information about each film with a content
based rating as input to the process of finding links to other similar films. There are several
ways to find links. We used a simple scoring mechanism. It then adds the film’s rating (either
predicted or user-given) to the score for each film element.
3. Analysis of Existing System and Limitation 51
Once it completes this process for all ratings, the agent calculates the average score for each
actor, director, and genre. This score indicates how much the user likes or dislikes each
element. The agent can then compute the predicted rating for each film. In a process similar to
that for finding links, the element’s average score is added to the film’s score. System
administrators who are configuring the recommender system can also assign weights to the
elements. The agent can then compute the predicted rating by dividing the film’s total score by
the number of elements used to calculate it. The agent can augment the list of films and
predicted ratings with any predictions that resulted from the initial collaborative-filtering
process but didn’t appear in the final prediction set (because of incomplete film information in
the database).
The agent also records the number of elements for each film as an indicator of the prediction’s
strength, again so it can use the information as a secondary sort field when it creates
recommendation lists.
In future work, to enable the assignment of different weightings to each filtering technique’s
results according to certain parameters. We can also reverse the order of algorithm. We want to
apply and evaluate our hybrid recommendation method to other domains and emerging
technologies.
10. Hybrid collaborative filtering algorithms using a mixture of experts
Author: Xiaoyuan Su1, Russell Greiner, Taghi M. Khoshgoftaar, Xingquan Zhu
Year: 2007
Summary:
Authors have proposed two hybrid CF algorithms, sequential mixture CF and joint mixture
CF. these approaches perform well in sparse data environment. A hybrid recommender
system combines CF and content-based techniques to overcome the limitations of either
recommender system and thereby improve recommendation performance. One shortcoming
of hybrid recommender systems is that the content information is not always available for
the reasons such as privacy protection.
Sequential mixture CF [30](SMCF) first uses the predictions from a TAN(tree augmented
naive Bayes network)-ELR(Extended Logistic regression) content-based predictor
(instead of NB) to fill in the missing values of the CF rating matrix to form a pseudo rating
3. Analysis of Existing System and Limitation 52
matrix, then predicts user ratings by using the Pearson CF algorithm instead of weighted
Pearson CF on the pseudo rating matrix. It is similar to content-boosted CF algorithm.
Fig 3.1: SMCF
Joint mixture CF (JMCF) [11]combines the predictions from three independent experts:
Pearson correlation-based CF, a pure TAN-ELR content-based predictor, and a pure TAN-
ELR model-based CF algorithm.
11. HYDRA-A hybrid recommendation system
Author: stephan spiegel, jérôme kunegis, fang li
Year: 2009
Summary:
Authors have combined CB and CF approach which utilize supplementary content feature
in order to improve the prediction accuracy. In Hydra, data normalization, feature
combination and matrix factorization are all preliminary steps to rating estimation.
Data normalization:
Subtractive normalization- few users sometime give higher rating than others, sometime
also items receive more positive feedback than other items. To compute accurate rating
prediction these global effect need to be removed from data.
Multiplicative normalization:
Subtractive normalization, the feature values would just be shifted instead of being
regularized. Therefore we make use of multiplicative normalization, which regularizes all
entries within a feature matrix f according their respective row and column length.
Feature combination:
The purpose is to identify those features and appropriate weights, which can improve the
prediction accuracy of our hybrid recommender system.
3. Analysis of Existing System and Limitation 53
Matrix factorization:
Matrix factorization techniques are used to reduce the dimension of the item space and/or to
retrieve latent relations between items of the observed dataset.
Hyb-svd-knn algorithm[13] (also referred to as hydra system) is able to raise prediction
accuracy by incorporating weighted user and item features. The pure SVD approach shows
the lowest computational effort; our hybrid method is about four times faster than
traditional collaborative filtering (KNN approach).
Here hybrid approach is special in that rating data as well as content information are joined
in a unified model, which leads to less parameters and more reasonable prediction results.
For the purpose of minimizing the runtime of designed hybrid recommender system as well
as to extract latent user and movie relations, factorize unified model by means of singular
value decomposition. The dimensionally reduced data can be employed to directly estimate
unknown ratings (pure SVD approach) or rather to accelerate collaborative filtering (SVD-
KNN as well as HYBSVD- KNN algorithm).
Future work, it would be interesting to apply this algorithm on different dataset from
different domain, because unlike content features might achieve even higher prediction
accuracy improvement.
12. A hybrid Recommendation Method with Reduced Data for large Scale
Author: Sang Hyun Choi, Young-Seon Jeong, and Myong K. Jeong
Year: 2010
Summary:
Authors have proposed HYRED algorithm which combine CF using the modified
Pearson’s binary correlation coefficients with CB filtering using the generalized distance-
to-boundary-based rating.
First, HYRED[32] proposes the concept of neighbourhood in CF to efficiently analyse the
transaction data. The use of the nearest and farthest neighbours of a target customer yields a
reduced dataset of useful information for solving the scalability problem. The organization
of the training dataset has been restricted to the items purchased by a target user and his or
her farthest neighbours so that the number of training datasets can be reduced considerably.
3. Analysis of Existing System and Limitation 54
At the testing step, we have found only items purchased by nearest neighbours and
predicted the score of each item. The use of fewer training and testing datasets enables us
not only to lessen the computing effort, but also to improve the performance of
recommendations.
The processes of filtering irrelevant data by using the neighbourhood concept of CF make it
possible to consider the items that are likely to be purchased by a target user. Second,
propose the generalized rating system based on the distance of an item to the decision
boundary of a classifier. In this concept, the item closer to the class of purchased items may
have a higher probability of being sold. The experiment shows that the DTB(Distance to
bound)-based rating improves the performance of recommendation than either pure CB or
CF. The algorithm has calculated the distance from alternative items to the ones purchased
by a target user, whose items have been selected from statistical classifiers. This selection
method used neighbourhood information and delivered better performance than was gained
with pure CF.
Finally, proposed a generalized hybrid recommendation algorithm by using a weighted
coefficient in which the DTB and CF methods are special cases of our generalized
algorithm. The weighting scheme makes this algorithm adequate for generalized
applications, and HYRED is flexible enough for application with any available datasets.
Moreover, HYRED, when weighting is properly valued, has yielded better results than pure
DTB, pure CF, and simple combined hybrid method.
13. Enhanching Accuracy of Recommender system through adapting the domain trends
Author: Fatih Aksel, Aysenur Birtürk
Year: 2010
Summary:
Authors proposed an adaptive method for hybrid recommender system, adarec, in which the
combination of algorithms are learned and dynamically updated from the results of
previous predictions.
Adarec, An Adaptive Hybrid Recommender System[33]:
It uses Hybrid recommendation systems which combine multiple algorithms and define a
switching behaviour (strategy) among them. This strategy decides which technique to
choose under what circumstances for a given prediction request. Adarec consists two parts:
3. Analysis of Existing System and Limitation 55
Recommendation Engine and Learning Module. Recommender Engine is responsible for
generating the predictions of items based on the previous user profiles and item contents.
The recommender generates the predictions by using its attached prediction strategy.
Learning Module handles the new prediction strategy creation upon the previous instances
and performance results of the prediction techniques on each learning cycle.
The learning module first tests the accuracy of the each predictor in the system. Than the
prediction strategy is redesigned by the learning module in order to improve proper use of
predictors. Adaptive prediction strategy improves its' prediction accuracy by learning better
when to use which predictors. The learning module adapts the hybrid recommender system
to the current characteristics of domain.
Research study shows that traditional static hybrid recommender systems suffer from
changing user preferences. In order to improve the recommendation performance, handle
domain drifts in our approach. The Learning Module re-designs its prediction (switching)
strategy according to the performance of prediction techniques based on user feedbacks. As
a result, the system adapts to the application domain, and the performance of
recommendation increases as more data are accumulated.
In future work, they have planned to further testing the learning module with various
heterogeneous datasets. It would be interesting to examine the different domains. In
experiments we fixed the used attributes for domain monitoring. It would be also
interesting to use dynamic attributes, which meansto use different attributes on different
iterations.
14. A Content enhanced approach for cold-start problem in collaborative filtering
Author: Dongting Sun, Cong Li and Zhigang Luo
Year: 2011
Approach:
Summary:
Author have proposed a hybrid algorithm by using both the rating and content information to
overcome cold-start (user and products without any rating) problem. This hybrid approach first
cluster items based on the rating matrix and then utilize the clustering results and item content
information to build a decision tree to associate the novel items with the existing ones. In cold
start problem, content information can help to bridge the gap between existing and new items,
3. Analysis of Existing System and Limitation 56
as well as between existing and new users by building relationships among them. Content
information can combine with collaborative in various ways.
The base of the algorithm is to find the similarity among items accurately. There are various
ways to compute the similarity, the most commonly used one is Pearson correlation.
IBCTAP Algorithm[34]-IBCTAP includes four main procedures: item clustering, decision tree
building, new item classifying and ratings predicting.
Item clustering: It reduces one large-dimensionality item-user space into a set of smaller
dimensionality spaces, with fewer items, less ratings, and often less users. The most popularly
used one is K-means Clustering. Pearson correlation coefficient is used to calculate similarity
between items. As output of this process, we can obtain the number of k clusters; items in each
cluster will be liked by the users with same tastes.
Decision tree building: In order to achieve an optimized decision tree we apply the most
commonly used standard, information gain, to decide which attribute is best to be chosen. The
algorithm first calculates the entropy of the whole data set. The algorithm calculates the
information gain for every attribute and chooses the one with the highest information gain.
After the root node has been decided, the algorithm creates two branches corresponding to true
or false. For each branch, the algorithm then determines if the branch can be divided further, if
can, the same method as above is used to determine which variable to use, if not, it has reached
a solid conclusion.
New items classifying: When a new item without any rating enters in the recommender
system, this algorithm captures the item content information immediately, and then follows
down the tree that we have completely trained in the tree building procedure. The decision tree
answers each question correctly and the new item will eventually arrive at a cluster.
Ratings predicting: In traditional collaborative filtering approach, it is hard to recommend
new items to users since the new item has not any past rated data. However, in this approach,
can recommend new items to users based on the hypothesis that the new item will be preferred
by the users who like the items in the cluster that the item has arrived in the classifying
procedure. Use MAE metrics for evaluating the accuracy of our prediction method. MAE
(Mean Absolute Error) has been widely used in evaluating the accuracy of a recommender
system by comparing predicted values with user-provided values.
3. Analysis of Existing System and Limitation 57
MAE=∑ | |
From experimental studies, it is observed that CF gives unexpected MAE value 3.1502 in
extreme cold situation. While algorithm gives 0.8251 MAE value which is four times more
improving than previous result.
Future work explores math function to measure the microcosm variety of the result curves. We
also find that items and users are symmetric under the view of rating data. If we invert the item-
user matrix direction and use the user content information to build decision tree, this model
may be applied to solve user-side cold start problem.
Table 3.1: Comparative study
Sr.
No
Year Title Work Future Work Author Publication
1 1997 Fab: Content-
based,
collaborative
recommendation
FAB
Meta-Level
Content
Based into
CF
Massive
Scale up
Dynamic
process
Balabanovic,
M., &
Shoham
Communicati
ons of the
ACM, 40(3),
66-72.
2 1998 Recommendatio
n as
classification:
using social and
content-based
information in
recommendation
Feature
Combination
Develope
unifying
Model
needed to
develop
instrument
s for
adapting a
general
recommen
der system
to a
specific
case
Adaption
could be
simplified
Chumki
Basu, Haym
Hirsh, and
William
Cohen
In
Proceedings
of the 1998
Workshop on
Recommender
Systems,
pages 11-15
3 1999 Combining
content-based
and
collaborative
filters in an
online
newspaper
Mixed
appraoch
Combining
Separate
Recommende
r
Linear
Combination
rating
Add
Demographics
technique
Accuracy in
prediction
Claypool,
M., Gokhale,
A., Miranda,
T.,
Murnikov,
P., Netes,
D., & Sartin,
M
ACM
SIGIR'99
Workshop on
Recommender
Systems:
Algorithms
and
Evaluation,
Berkeley, CA.
4 1999 Combining
collaborative
filtering with
Add content
based
charachteristi
Further
Combination
user and
Good, N.,
Schafer, J.
B., Konstan,
Proceedings
of the
Sixteenth
3. Analysis of Existing System and Limitation 58
personal agents
for better
recommendation
cs to
collorative
methods
agents
Scalability
J. A.,
Borchers,
A., Sarwar,
B.,
Herlocker,
J., & Riedl,
J.
National
Conference on
Artificial
Intelligence,
Orlando, FL,
pp. 439-446.
5 2000 PTV: Intelligent
Personalised TV
Guides
Mixed
Approach
WAP
version of
PTV
launched
Paul Cotter
& Barry
Smyth
IAAI-00
Proceedings.
Copyright ©
2000, AAAI
(www.aaai.or
g)
6 2001 “Probabilistic
models for
unified
collaborative
and
contentbased
recommendation
in sparse-data
environments
Unified
Model based
approach
reduce the
need for
multiple
restarts
Used where
it is actually
followed by
user
A. Popescul,
L. H. Ungar,
D. M.
Pennock,
and S.
Lawrence
Proceedings
of the 17th
Conference in
Uncertainty in
Artificial
Intelligence
(UAI ’01), pp.
437–444
7 2002 Content-boosted
collaborative
filtering for
improved
recommendation
s.
content
within a
collaborative
framework
improvem
ents in
collaborati
vefiltering
or content-
based
recommen
ding
Work on
accurate
prediction
Melville, P.,
Mooney, R.
J., &
Nagarajan,
R.
Proceeding
s of the
Eighteenth
National
Conference
on
Artificial
Intelligence
(pp. 187-
192).
Menlo
Park, CA /
Cambridge,
MA: AAAI
Press / MIT
Press.
8 2003 Clustering
Approach for
Hybrid
Recommender
System,
Work on cold
start problem
by clustering
approach
---- Li, Q. &
Kim, B. M
Proc. of the
IEEE/ WIC
International
Conference on
Web
Intelligence,
pp. 33-38,
9 2006 CinemaScreen
Recommender
Agent:Combinin
g Collaborative
and Content-
Based Filtering
Content
based
filtering on
CF result
reverse the
order of
algorithm
Apply on
other
James Salter
and Nick
Antonopoulo
s
Intelligent
Systems,
IEEE,
Volume: 21 ,
Issue: 1
3. Analysis of Existing System and Limitation 59
domain
10 2007 Hybrid
collaborative
filtering
algorithms using
a mixture of
experts
Incorporatein
g CF and
Content
Based
Features
----- X. Su, R.
Greiner, T.
M.
Khoshgoftaa
r, and X.
Zhu
Proceedings
of the
IEEE/WIC/A
CM
International
Conference
onWeb
Intelligence
(WI ’07), pp.
645–649,
Silicon
Valley,
Calif, USA
11 2009 HYDRA-A
hybrid
recommendation
system
Minimize
runtime
Choose
content
accurate
content
features for
accuracy
Apply on
different
attribute
domain and
check
accuracy
Stephan
S.,Jerome
K,Fang L.
CNIMK’09
ACM
12 2010 A hybrid
Recommendatio
n Method with
Reduced Data
for large Scale
Accuracy is
improve here
over full data
set
Apply for
other
recommendati
on domain
Sang hyun
choi,young-
seon,mayon
g
IEEE
transaction on
systems ,man
and
cybernetics-
Part
C:application
and review
vol 40 no 5
13 2010 Enhanching
Accuracy of
Recommender
system through
adapting the
domain trends
Dynamically
update the
result
Effectivness
accuracy of
other machine
learning
techniques,
Dynamic
attributes
Faith
a.,Aysenur
B.
PRSAT 2010
held in
conjunction
with RecSys
2010
14 2011 A Content
enhanced
approach for
cold-start
problem in
collaborative
filtering
Solve cold start
problem
Combine
algorithm more
accurately
DongtingS.,
Cong
L.,Zhigang
L.
IEEE
3. Analysis of Existing System and Limitation 60
3.2 RELATED WORK
Initially Fab maintains user profiles of interest in web pages using content-based techniques, and
uses CF techniques to identify profiles with similar tastes. It can then recommend documents across
user profiles[35].
In next approach, simply both CB and CF methods produce separate result and then combine their
prediction[26].
In another approach[25], treat recommending as a classification task. It uses both user ratings and
contents features to produce recommendations .They use Ripper, a rule induction system, to learn a
function that takes a user and movie and predicts whether the movie will be liked or disliked. They
combine collaborative and content information, by creating features.
In another approach [36], the term-document matrix is multiplied with the user-ratings matrix to
produce a contentprofile matrix. Using Latent Semantic Indexing, a rank k approximation of the
content-profile matrix is computed. Term vectors of the user’s relevant documents are averaged to
produce a user’s profile. Now, new documents are ranked against each user’s profile in the LSI
space.
In another approach[37], each user-profile is represented by a vector of weighted words derived from
positive training examples using the Winnow algorithm. Predictions are made by applying CF
directly to the matrix of user-profiles (as opposed to the user-ratings matrix).
In this approach [38], it uses collaborative filtering along with a number of personalized information
filtering agents. Predictions for a user are made by applying CF on the set of other users and the
active user’s personalized agents. Our method differs from this by also using CF on the personalized
agents of the other users.
In this approach [39], implemented a set of knowledge-based “filterbots” as artificial users using
certain criteria. A straightforward example of a filterbot is a genrebot, which bases its opinion solely
on the genre of the item, for example, a “jazzbot” would give a full mark to a CD simply because it is
in the jazz category, while it would give a low score to any other CD in the database.
In this approach [40], it uses the prediction from the CF system as the input to a content-based
recommender.
3. Analysis of Existing System and Limitation 61
It [41] proposes a Bayesian mixed-effects model that integrates user ratings, user, and item features
in a single unified framework.
In Content – Boosted CF [42]. Uses na¨ıve Bayes as the content classifier, it then fills in the missing
values of the rating matrix with the predictions of the content predictor to form a pseudo rating
matrix, in which observed ratings are kept untouched and missing ratings are replaced by the
predictions of a content predictor. It then makes predictions over the resulting pseudo ratings matrix
using a weighted Pearson correlation-based CF algorithm, which gives a higher weight for the item
that more users rated, and gives a higher weight for the active user.
In another approach[30], used TANELR as the content-predictor and directly applied the Pearson
correlation-based CF instead of a weighted one on the pseudo rating matrix to make predictions, and
they achieved improved CF performance in terms of MAE.
In another approach, [43] propose a Bayesian preference model that statistically integrates several
types of information useful for making recommendations, such as user preferences, user and item
features, and expert evaluations. They use Markov chain Monte Carlo (MCMC) methods for
sampling based inference, which involve sampling parameter estimation from the full conditional
distribution of parameters. They achieved better performance than pure collaborative filtering.
3.3 HYBRID TECHNIQUE – CBCF
3.3.1 Introduction
Melville et al. Proposed a content-boosted collaborative filtering algorithm (CBCF) to
overcome the shortcomings of content-based and collaborative filtering algorithms when used
individually [42]. By using a content-boosted collaborative filtering approach, the authors
solved the sparsity and cold-start problems of collaborative filtering algorithms that were
discussed in the previous chapter. They used the content information to seed the user-ratings
matrix to solve the two problems. The shortcomings of content-based methods of not finding
serendipitous recommendations were also solved using social information coming from the
data set of a social network.
The CBCF algorithm was tested on movie recommendations, A pure content-based predictor
was used to learn a user's rating and predict the ratings for the unrated movies. The pure
content-based predictor uses a naive Bayesian classifier to learn and predict user ratings. The
predicted ratings along with the user ratings of each user are referred to as the pseudo user-
3. Analysis of Existing System and Limitation 62
ratings vector. This pseudo user-ratings matrix is used by a pure collaborative filtering
algorithm that uses a neighbourhood-based algorithm to find a subset of users who are similar
to an active user. A set of neighbours are chosen who have the highest similarity to the active
user measured by the Pearson correlation coefficient. At the end compute a prediction from
a weighted combination of the selected neighbours’ ratings.
3.3.2 Discussion
In this section we explain how content-boosted collaborative filtering overcomes some of the
shortcomings of pure CF[44].
Overcoming the First-Rater Problem
Pure CF suffers from first rater problem/New user-item problem. First rater problem means an
item which does not have any past rating data. However, here this kind of prediction is
possible with the help of content based predictor. If the neighbors of the active user are highly
correlated to it, then their CB predictions should also be very relevant to the user. This is
particularly true if neighbors have rated many more items than the active user; because their
CB predictions are likely to be more accurate than the active user’s. In this way, CBCF solves
the first-rater problem, and produces even better predictions than the content-based predictor.
Tackling sparsity
In CBCF, since we use a pseudo ratings matrix, which is a full matrix, we eliminate the root
of the sparsity problem. Pseudo user-ratings vectors contain ratings for all items; and hence all
users will be considered as potential neighbors. This increases the chances of finding similar
users. Thus the sparsity of the user-ratings matrix affects CBCF to a smaller degree than CF.
Finding better neighbours
A crucial step in cf is the selection of a neighborhood. The neighbors of the active user
entirely determine his predictions. It is therefore critical to select neighbors who are most
similar to the active user. In pure cf, the neighborhood comprises of the users that have the
best n correlations with the active user. The similarity between users is only determined by
the ratings given to co-rated items; so items that have not been rated by both users are
ignored. However, in cbcf, the similarity is based on the ratings contained in the pseudo user-
ratings vectors; so users do not need to have a high overlap of co-rated items to be considered
similar.
3. Analysis of Existing System and Limitation 63
3.3.3 Comparison of Naïve Bayes with other Content Based technique
One technique adapted from IR is the assignment of weights to keywords. The commonly
used approach to specify weights to keywords is the term frequency-inverse document
frequency (TF-IDF). Other techniques used for content-based Recommendations include
probabilistic models, such as Bayesian classifiers, and machine learning techniques like
artificial neural networks. These approaches generate predictions by learning the underlying
model with statistical analysis and machine learning techniques [45]. This thesis uses content-
based method with Naïve Bayes.
Traditional methods (Heuristics) based on information retrieval while other methods are
calculating utility prediction. In our thesis, we require prediction so here naïve Bayes method
is selected for content based predictor. Because of simplicity and effectiveness, also Naïve
Bayes classifiers are often used in text classification applications and experiments. Here we
compare the Bayesian classifier to several standard machine learning algorithms and present
experimental evidence that the Bayesian classifier performs at least as well as these
computationally more intensive alternatives.[46]
Nearest neighbor
The nearest neighbor algorithm operates by storing all examples in the training set. To
classify an unseen instance, it assigns it to the class of the most similar example. Since all of
the features we use are binary features, the most similar example is the one that has the most
feature values in common with a test example.
Decision trees
Decision tree learners such as ID3 build a decision tree by recursively partitioning examples
into subgroups until those subgroups contain examples of a single class. A partition is
formed by a test on some attribute (e.g., is the feature database equal to 0). ID3 selects the
test that provides the highest gain in information content.
Neural nets
We used two approaches to learning with neural nets. In the perceptron approach, there are
no hidden units and the single output unit is trained with the delta rule .The perceptron is
limited to learning linearly separable functions. We also use multi-layer networks trained
with error back propagation. We used 12 hidden units in our experiments.
3. Analysis of Existing System and Limitation 64
Table 3.2 [46]: Average accuracy of the classification algorithm
So here Bayes Classifier performs consistently well on most domains. It is also very fast for
both learning and predicting.
3.3.4 Comparison of Pearson Co-relation with other similarity measures in CF
There is a variety of similarity metrics available [47]. Some of the most commonly used
measures include Pearson correlation coefficient, cosine measure, distance measure and
jaccard coefficient.
Distance measure can be used is the distance. Distance between data objects is sum of the
distances of each attribute of the data objects (i.e. Euclidean Distance).
Another method the machines can use to determine the similarity between data objects is by
measuring how the attributes of both data objects change with respect to the variation of the
mean value for the attributes. This method of determining the similarity is the Pearson
Correlation coefficient.
There may be cases when the data objects are not simply a group of numbers, but perhaps a
Boolean value. To represent the similarity to a machine, finding the ratio between the
numbers of matching attributes to the total number of attributes is a better metric, which is the
case with the Jaccard Coefficient.
The cosine similarity may be used and an example of this metric being used is with document
comparison. By using the word frequencies for each document, the normalized dot product of
the frequencies can be used as a measure of similarity.
3. Analysis of Existing System and Limitation 65
Table 3.3: Various Similarity Measures
Similarity
Measure
Used Remarks
Eucledian
Distance
To calculate
distance between two points
-
Cosine
Similarity
To determine similarity
between two documents
Since there are more words that are
incommon between two documents,
it is useless to use the other methods
of calculating similarities (namely
the Euclidean Distance and the
Pearson Correlation Coefficient
discussed earlier)
Jaccard
Coefficient
each attribute is binary such
that each bit represents the
absence of presence of a
characteristic,
-
Pearson
correlation
coefficient
TO measures how highly
correlated are two variables
and is measured from -1 to
+1
Used in User Based CF
Pearson correlation coefficient:
Here in cbcf, full matrix is provided as input to cf. It contains rating of movie given by user.
In cf process, pure cf algorithm that uses a neighborhood-based algorithm to find a subset of
users, who are similar to an active user. A set of neighbors are chosen who have the highest
similarity to the active user. So to measure highly correlated users, pearson correlation
coefficient a best option.
There are several benefits to using this type of metric. The first is that the accuracy of the
score increases when data is not normalized. As a result, this metric can be used when
quantities (i.e. Scores) varies. Another benefit is that the Pearson correlation score can correct
for any scaling within an attribute.
3. Analysis of Existing System and Limitation 66
3.4 Evaluation measure of rs
Evaluation still presents several challenges and problems, summarized here[48]:
(1) Coverage
This corresponds to the percentage of items the system is able to recommend
(2) Prediction accuracy
This measures the difference between the rating the system predicts and the real rating.
The most popular of this kind of metric is the mean absolute error (MAE).
MAE=∑ | |
Where n is the total number of ratings over all users, pi, j is the predicted rating for user i on
item j, and ri, j is the actual rating [49].
Other related metrics, such as mean squared error (mse), root mean squared error (rmse):
RMSE=√
∑
Where n is the total number of ratings over all users, pi, j is the predicted rating for user i on
item j, and ri, j is the actual rating again. RMSE amplifies the contributions of the absolute
errors between the predictions and the true values.[49]
3) Classification accuracy and Rank Accuracy
This measures how well the system differentiates good items from bad ones. Examples of well-
known metrics of this type are Precision, Recall and ROC [47]. These metrics are appropriate
for the find good items task, especially when the preferences of the users are binary.
Table 3.4: confusion matrix
Reality All
recommended
items All good
Items
Good Bad
Prediction Related
Good
True Positive False Positive
Related Bad False Negative True Negative
3. Analysis of Existing System and Limitation 67
Precision measure the degree of accuracy of recommendation produced by the system.
Precision=
=
Recall measure the degree of relevant recommendation to the total number of
recommendation.
Recall=
=
ROC is used a plot of the system sensitivity and (1-specificity),where sensitivity is the
probability of a randomly selected good item being recommended by the system and
specificity is the probability of a randomly selected item being refused by the system.
Rank accuracy measures the ability of the system to sort the recommended items like the user
would have done. In many cases, this kind of metrics is too sensitive given they ask the system
to recommend the best items when, in practice, it would suffice to recommend good items and
not necessarily the best.
3.5 DATASET [44]
3.5.1 Overview
Study experiments are evaluated in movie lens dataset which is provided by the Compaq
Systems Research Center. This dataset contains 7,893 randomly selected users and 1,461
movies for which content were available from the Internet Movie Database (IMDB). The
reduced dataset has 299,997 ratings for 1,408 movies and average number of votes per user is
approximately 38. Minimum value of rating is zero and maximum rating value is 5.
Actually, the important point is whether movie lens dataset will be dense or sparse when the
missing data prediction process is handled. Initial sparsity of movie lens dataset is 97.4 %.
In order to evaluate the prediction mechanism of system, cross validation method was used
and among the various cross validation methods, the holdout method was preferred.
Following this method, the data set was separated into two sets, called the training set and the
testing set. We represent the content information of every movie as a set of slots (features).
3. Analysis of Existing System and Limitation 68
Each slot is represented simply as a bag of words. The slots we use for the each- Movie
dataset are: movie title, director, cast, genre, plot summary, plot keywords, user comments,
external reviews, newsgroup reviews, and awards.
3.5.2 Result Analysis
This section discusses the results that were obtained with the experiments set up.
There are four methods (pure CB, pure CF, naïve hybrid and CBCF) apply on movie lens
dataset and give various MAE and ROC-4 value.
The naive hybrid approach takes the average of the ratings generated by the pure content-
based predictor and the pure CF predictor. Ten percent of the users were randomly selected to
be the test users. From each user in the test set, ratings for 25% of items were withheld.
Predictions were computed for the withheld items using each of the different predictors.
Table 3.5 [44]: Summary of Result
CBCF was significantly better than the other algorithms at both MAE(0.956) and ROC-4
(0.7716).On the MAE metric, CBCF performs 9.2% better than pure CB, 4% better than pure
CF and 4.9% better than the naive hybrid.
On the ROC-4, metric CBCF performs 5.4% better than pure CB, 4.6% better than pure CF and
9.7% better than the naive hybrid.
3.6 LIMITATION OF EXISTING SYSTEM
These systems can still suffer from scalability problems as the number of items and users
increases exponentially [42]. Web services in particular suffer from producing
recommendations of millions of items to millions of users. The time and computational power
can even limit the performance of the best hybrid systems.
Although CBCF performs consistently better than pure CF, the difference in performance is
not very large (4%).[44]