fake review detection
TRANSCRIPT
FAKE REVIEW DETECTION What are they saying about you ? Are they real …
Guided By : Dr. Animesh Mukherjee
ONLINE REVIEW
• Captures testimonials of “real” people (unlike advertisements).
• Shapes decision making of customers.
• Positive reviews – financial gains and fame for business.
• Deceptive opinion spamming – to promote or discredit some target product
and services.[1]
• Opinion Spammers admitted for being paid to write fake reviews.(Kost, 2012)
• Yelp.com – “Sting operation” : publically shame business who buy fake
reviews.[2]
[1] Jindal and Liu 2008 : http://www.cs.uic.edu/~liub/FBS/opinion-spam-WSDM-08.pdf
[2] Yelp official blog : http://officialblog.yelp.com/2013/05/how-yelp-protects-consumers-from-fake-reviews.html
Consumer Alerts !!
• Amazon Mechanical Turk – crowd sourced online workers(turkers) to write fake
reviews ($1 per review) portraying 20 hotels of Chicago in positive light.[1]
• 400 fake positive reviews collected and 400 non-fake reviews on same 20 hotels
using Tripadvisor.com
• Yelp’s filtered and unfiltered reviews collected to understand working of Yelp
review classification algorithm.
• Approach : Linguistic n-gram features and some supervised learning method.[2]
[1] Amazon Mechanical Turk : https://www.mturk.com/
[2] Ott et al. 2011 : https://www.cs.cornell.edu/courses/CS4740/2012sp/lectures/op_spamACL2011.pdf
Dataset Collection and Approach for Analysis
Linguistic Approach : Results
• Using only Bigram features :
accuracy of 89.6% on AMT data.[1]
• Using same n-gram features :
accuracy of 67.8% on Yelp Data.[2]
• Table 1: Class distribution of Yelp
data is skewed – imbalanced data
produces poor model.(Chawla et al,
2004)
• Good model for imbalanced data –
under sampling. (Drummond and
Holte, 2003)
[1] Ott et al. 2011 : https://www.cs.cornell.edu/courses/CS4740/2012sp/lectures/op_spamACL2011.pdf
[2] A Mukherjee - 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
Linguistic Approach Results : Explained
• For AMT data, word distribution of fake and non-fake reviews are very different,
which explains the high accuracy using n-gram.
• Reason for the different word distribution : Domain knowledge absence and little
gain in writing fake reviews($1 per review).
• Poor performance of n-grams for Yelp data because spammers according to
Yelp filter used very similar language in fake review as non-fake review :
linguistically very similar.[1]
• Inefficiency in linguistic in detecting fake reviews filtered by Yelp encourages to
do behavioral study of reviews.
[1] Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
))(/)((2log)()||( iNiFiFNFKLi
Information Theoretic Analysis
• To explain huge accuracy difference : analysis of word distribution of AMT and
Yelp data.
• Good-Turing smoothed unigram language models.
• Computation of word distribution difference across fake & non-fake reviews :
Kullback-Leibler(KL) Divergence :
where F(i) and N(i) are respective probabilities of word i in fake and non-fake reviews.[1]
• Here KL(F||N) gives quantitative estimate of how much fake reviews linguistically
differ from non-fake reviews
[1] Kullback Leibler Divergence : http://www.cs.buap.mx/~dpinto/research/CICLing07_1/Pinto06c/node2.html
Information Theoretic Analysis (Cont..)
• KL Divergence is Asymmetric :
• We have,
• For AMT data : and
• For Yelp data : and
• [1]
FNKLNFKL ||||
FNKLNFKL |||| 0KL
FNKLNFKL |||| 0KL
[1] Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
FNKLNFKLKL ||||
Information Theoretic Analysis (Cont..)
• KL(F||N) Definition implies that words having higher probability in F and very low
probability in N contribute most to KL-Divergence.
• To study word-wise contribution to ∆KL, word ∆KL calculated as :
where,
• Contribution of top k words to ∆KL for k= 200 and k=300
)||()||( iiwordiiword
i
Word FNKLNFKLKL
)(
)(log)||( 2
iN
iFiFNFKL iiword
Turkers didn’t do a good job at “Faking”!
Word-wise Difference of KL-
Divergence across top 200 words.
Equally Dense : |E| = |G|
• Symmetric Distribution of for top k-words for AMT
data implies the existence of two set of words:
1. set of words E, appearing more in fake reviews than
in non-fake, Ɐi € E, F(i) > N(i) resulting in
2. set of words G, appearing more in non-fake reviews
than in fake, Ɐi € G, N(i) > F(i) resulting in
i
WordKL
0 Ei
WordKL
0 Ei
WordKL
• Additionally top k=200 words only contribute 20% to ∆KL for the AMT data.
There are many words in AMT data having higher probabilities in fake than non-
fake and vice-versa.
• This implies fake and non-fake reviews in AMT data consist of words with very
different frequencies. Turkers didn’t do a good job in faking.[1]
[1] Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
Yelp Spammers are Smart but Overdid “Faking”!
• Yelp fake review data shows KL(F||N) is much larger than KL(N||F) and ∆KL>1.
• Given graphs from b-e shows among top 200 words which contribute to major
(=70%) most words have and only a few have .
Word-wise Difference of KL-Divergence across top 200 words.[1]
0 i
WordKL 0 i
WordKL
[1] Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
Yelp Spammers are Smart but Overdid “Faking”! (Cont..)
• Consider A be set of top words contributing most to ∆KL. We partition A as:
where and
• The curve above y=0 is dense and below it is sparse which implies
• clearly indicates that there exists specific words which contribute
most to ∆KL by appearing in fake reviews with much higher frequencies than
in non-fake reviews.
• Spammers made smart effort to ensure that their fake reviews have most
words that also appear in non-fake reviews to sound convincing.
))()(,.,(0| iFiNAiieKLiA Ni
Word
N
))()(,.,(0| iNiFAiieKLiA Fi
Word
F NF AAA NF AA
NF AA
NF AA
Yelp Spammers are Smart but Overdid “Faking”! (Cont..)
• While making their reviews sound convincing, psychologically they
happened to OVERUSE come words resulting higher frequency of certain
words in fake review than non-fake reviews.
• A quick lookup yields {us, price, stay, feel, deal, comfort} in hotel domain and
{options, went, seat, helpful, overall, serve, amount etc.} in restaurants.
• Prior personality work shown that deception/lying usually involves more use
of personal pronouns(eg. us) and associated action(eg. Went, feel) towards
specific targets(eg. option, price, stay) with objective of incorrect
projection(lying or faking).[1]
• Spammers caught by Yelp left behind linguistic footprints which can be
caught by precise behavioral study.
[1] Newman et al , 2003 : http://www.communicationcache.com/uploads/1/0/8/8/10887248/lying_words-_predicting_deception_from_linguistic_styles.pdf
Spamming Behavior Analysis
1. Maximum Number of Reviews(MNR) : Writing too many reviews in a day is
abnormal.
2. Percentage of Positive Reviews(PR) : Deception words in fake reviews
indicates projection in positive light. CDF of positive(4-5 stars) reviews
among all reviews is plotted to illustrate analysis.
3. Review Length(RL) : While writing fake experiences, there is probably not
much to write and also spammer not want to spend too much time in it.
4. Maximum Content Similarity(MCS) : To examine if some posted reviews are
similar to previous reviews, we computed cosine similarity between two
reviews of a reviewer. Non-spammers mostly write new contents.
Maximum Number of Reviews Percentage of Positive Reviews
Review Length Maximum Content Similarity
Spamming Behavior Analysis CDFs
[1] Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
Challenges with Supervised Evaluation
• Very difficult to find gold-standard data of fake and non-fake reviews for
model building – too difficult to manually recognize/label fake/non-fake
reviews by mere reading.
• Duplicate and near duplicate assumed to be fake, which is unreliable.
• Usage of manually labeled dataset – reliability issues because it have been
shown that accuracy of human labeling of fake reviews is very poor.[1]
• AMT crowdsourcing fake reviews by paying are fake yet they don’t reflect the
dynamics of fake reviews in commercial website.[2]
• This lack of labeled data, motivates to look after the unsupervised methods of
classification.[3]
[1] Ott M, Choi, Y, Cardie, C. and Hancock, J.T. 2011. Finding Deceptive Opinion Spam by Any Stretch of the Imagination.
[2] Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-submit.pdf
[3] Mukherjee et al : http://delivery.acm.org/10.1145/2490000/2487580/p632-mukherjee.pdf
Unsupervised Evaluation Model
• Since human labeling for supervised learning is difficult, problem was
proposed by modeling spamicity (degree of spamming) as latent with other
observed behavioral features.
• Unsupervised model – Author Spamicity Model (ASM) proposed.[1]
• Taken a fully Bayesian approach and formulated opinion spam detection as
clustering problem.
• Opinion spammers have different behavioral distributions than non-spammers.
• Causes distributional divergence between latent population distributions of two
clusters: spammers and non-spammers.[1]
• Model inference results in learning the population distributions of two clusters.
[1] Mukherjee et al : http://delivery.acm.org/10.1145/2490000/2487580/p632-mukherjee.pdf
• Formulates spam detection as an unsupervised clustering problem in
Bayesian setting.
• Belongs to class of Generative models for clustering based on set of
observed features.
• Models spamicity 𝑠𝑎(in range[0, 1]) of an author a; and spam label 𝜋𝑟 of a
review which is the class variable reflecting the cluster membership (two
cluster K=2, spam and non-spam).[1]
• Each author/reviewer and respectively each review has a set of observed
features(behavioral clues).
• Certain characteristics of abnormal behavior defined which likely to link
with spamming and thus can be exploited in model for learning spam and
non-spam clusters. [1]
Author Spamicity Model
[1] Mukherjee et al : http://delivery.acm.org/10.1145/2490000/2487580/p632-mukherjee.pdf
Author Features
• It have value in range [0, 1] and value close to 1 indicates spamming.
1. Content Similarity : Crafting new review every time is time consuming, spammers
likely to copy reviews across similar product. Choose maximum similarity to
capture worst spamming behavior.[1]
2. Maximum Number of Reviews : Posting many reviews a day is also abnormal.
3. Reviewing Burstiness : Spammers are usually not longtime members of site.
Defined over an activity window(first and last review posting date). If posted over
reasonably long timeframe, it probably a normal activity but all review posted
within short burst likely to be spam.[2]
4. Ratio of First Review : People mostly rely on early reviews and spamming early
impacts hugely on sales. Spammers try to be among first reviewers. [1]
[1] Mukherjee et al : http://delivery.acm.org/10.1145/2490000/2487580/p632-mukherjee.pdf
[2] Mukherjee, A., Liu, B. and Glance, N. 2012. Spotting Fake Reviewer Groups in Consumer Reviews. WWW (2012).
Review Features
• It have five binary review features. Value 1 indicates spamming else 0 is non-spamming.
1. Duplicate /Near Duplicate Reviews: Spammers often post multiple reviews which are
duplicate /near duplicate on same product to boost ratings.
2. Extreme Rating: Spammers mostly like to give extreme ratings(1 or 5) in order to boost
ratings to demote or promote products.
3. Rating Deviation: Spammers usually involve in wrong projection in either positive or
negative and it deviates from average ratings given by other reviewers.
4. Early Time Frame: Early review can greatly impact people’s sentiments on a product.
5. Rating Abuse: Multiple ratings on same product are unusual. Similar to DUP but focuses
on rating dimensions rather than content.
CONCLUSIONS
• We presented an in-depth investigation of nature of fake reviews in Commercial
settings of Yelp.com.
• Our study shows linguistics methods of (Ott et al., 2011) and its high accuracy in
AMT data.
• We presented a behavioral study of spammers for real-life fake reviews.
• We presented a brief introduction to n-gram language model.
• We presented challenges with the supervised evaluation and gave idea about the
unsupervised approach of evaluation.
• We presented a brief introduction to unsupervised Author Spamicity Model (ASM).
REFERENCES
1. Jindal and Liu 2008 : http://www.cs.uic.edu/~liub/FBS/opinion-spam-WSDM-08.pdf
2. Yelp official blog : http://officialblog.yelp.com/2013/05/how-yelp-protects-consumers-from-fake-
reviews.html
3. MIT N-gram Language Model Tutorial:
http://web.mit.edu/6.863/www/fall2012/readings/ngrampages.pdf
4. Amazon Mechanical Turk : https://www.mturk.com/
5. Ott et al. 2011 :
https://www.cs.cornell.edu/courses/CS4740/2012sp/lectures/op_spamACL2011.pdf
6. Mukherjee et al , 2013 : http://www2.cs.uh.edu/~arjun/papers/ICWSM-Spam_final_camera-
submit.pdf
7. Kullback Leibler Divergence :
http://www.cs.buap.mx/~dpinto/research/CICLing07_1/Pinto06c/node2.html
8. Newman et al , 2003 :
http://www.communicationcache.com/uploads/1/0/8/8/10887248/lying_words-
_predicting_deception_from_linguistic_styles.pdf
9. Mukherjee, A., Liu, B. and Glance, N. 2012. Spotting Fake Reviewer Groups in Consumer
Reviews. WWW (2012).
10. Mukherjee et al : http://delivery.acm.org/10.1145/2490000/2487580/p632-mukherjee.pdf
11. Ott M, Choi, Y, Cardie, C. and Hancock, J.T. 2011. Finding Deceptive Opinion Spam by Any
Stretch of the Imagination.
Any Question ?
Thank You !!