Review Spam Detection via Temporal Pattern Discovery Sihong Xie, Guan Wang, Shuyang Lin, Philip S. Yu Department of Computer Science University of Illinois at Chicago Give some examples of spam reviews Also give brief descriptions What’s review spams • Created on review websites in order to create positive impressions for bad products/stores, and make profit out of misled customers • They are harmful: lead to poor customer experience, ruin reputation of good stores • Guidelines to spot fake reviews for http://consumerist.com/2010/04/how-you-spot-fake-online-reviews.html human, but it is hard for machines ([1]) Human friendly clues of spams • Language features [5] Too hard for machines: Involving natural language processing 1.All praises 2.say nothing about the product 3.Red flag words 4.mention the name a lot Machine friendly clues of • similar reviewsspams (texts and ratings) on one product / product group in a short time [1,2] Machine friendly clues of spams • Group of spammers: • Wrote reviews together frequently on the same set of products / stores [3] Reviewer 1 Reviewer 2 Reviewer 3 How to play the spamming game Player 1 Player 2 Detection systems • Duplicated reviews: Easy: shingling detection is easy [3] detection based on statistics of reviewer • • two reviews are almost the same Group spamming: a group of reviewers frequently write reviews together. Other kinds: targeting on the same product, similar texts/ratings by one id. Spammers More sophisticated writings Use different reviewer ids Use different reviewer ids Failed machine friendly clues of spams • If these reviews were posted by the same id, it would have been easy to detect Failed machine friendlyclues of spams • Same id wrote multiple reviews, making it easy to be detected. Smart spammers would avoid this Reviewer 1 Reviewer 2 Reviewer 3 Spammers like singleton spam • Strong motivations to have singleton spams: 1. Need to boost the rating in a short time 2. Need to avoid being caught 3. Post reviews with high rating under different names in a short time Singleton reviews Registration Singleton 0 A physical person can register many reviewer ids + non-singleton 0 Normal reviewer + Spammer Reviewer id Store Each reviewer id contributes only one review for one store only Facts of Singleton reviews • Constitute a large portion of all the reviews • • Over 90% of the reviews are singleton reviews in this paper; similar situations in another dataset [4] More influential, more harmful The challenges Traditional clues shortcomings Review features (bag of words, ratings, brand names reference) [4] Hard for human, not to mention machines Reviewer features (rating behaviors) [1] Poor if one wrote only one review Product/Store features[4] Review/reviewer/store reinforcements Group spamming [2,3] Tell little about individual reviews Fails on large number of spam reviews with consistent ratings No applicable on singleton reviews Finds suspicious hotels, Singleton reviews detection [7]* can’t find individual singleton * [7] is a supervised method, and we have contrasting conclusion with theirs spam The proposed method • Recall the motivations of singleton reviews: boost the ratings in a short time and avoid being caught • The results: in a short time, many reviewers wrote only one review with a very high rating • The correlations between rating and volume of (singleton) reviews is the key feature of singleton review spamming Detected burst of singleton spams average rating number of reviews ratio of singleton reviews a suspicious time window The algorithm 1. For each store do A. split the whole period into small time windows B. compute avg rating, total number of reviews, percentage of singleton reviews in each window C. form a three dimension time series D. detect windows with correlated burst patterns E. for each detected window, repeat step A.-D. until window size becomes too small The algorithm 1 sorted by posting time; divided into groups 3 average rating: 2 review volume: 3 SR volume: 1/3 2 5 Multi-dimensional time series average rating: 4.6 review volume: 5 SR volume: 5/5 4 5 average rating: 2 review volume: 3 SR volume: 3/3 5 4 1 the correlated burst 3 2 • • • • • Dataset A snapshot of a review website* 408,469 reviews, 343,629 reviewers 310,499 reviewers (> 90%) wrote only one review 76% reviews are singleton reviews Focus on top 53 stores with over 1,000 reviews # reviewers * www.resellerratings.com # reviews Experimental results • • 29 stores are regarded as suspicious by least 2 out of 3 human evaluators. The proposed algorithm labeled 39 stores as suspicious. ( recall = 75.86%, precision = 61.11%) Case studies time window size = 30 days correlated bursts detected Period with the detected correlated burst enlarged time window size = 15 days pin-point the exact time and shape of the bursts Volume of reviews: 154 57 83% Ratio of SRs: 61% rating: 4.56 4.79 • • • Case studies (cont’) Text features: ratio of reviews talking about “customer service/support” Hurry reviewers: wrote only one review at the same time of ID registration Human validation: read the reviews and found a reviewer disclosed being solicited for a 5 star review Most of the later reviews are written by “Hurry Reviewers” more than 80% of the singleton reviews are related to “customer service” more than 80% of the singleton reviews are related to “customer service” References 1. Detecting Product Review Spammers using Rating Behaviors 2. Finding Unusual Review Patterns Using Unexpected Rules 3. Spotting Fake Reviewer Groups in Consumer Reviews 4. Opinion spam and analysis 5. Finding Deceptive Opinion Spam by Any Stretch of the Imagination 6. Review Graph based Online Store Review Spammer Detection 7. Merging Multiple Criteria to Identify Suspicious Reviews the end Examples All praises Posted in a short time say nothing about the product similar ratings Red flag words mention the name a lot Need to set up the animations for all these elements Types of review spams • Duplicate (easy, string matching) • Advertisements • Other easy-to-detect spams (all symbols, numbers, empty, etc.) • untruthful (very hard, need machines to understand the intentions of the reviews) Feature based methods • Paradigms: • Define features, training set and pick a classifier • Keys to success: good features, large training data, and powerful classifier Common but not fully investigated • Previous methods can not catch them • Each reviewer id has only one review • Many features used by previous methods simply become meaningless Traditional methods (cont.) • [6] uses a graph to describe reinforcement relationships between entities: good / bad reviews influence their authors, who in turn influence the stores, which in turn influence its reviews. reviews reviewers stores Traditional methods (cont.) The store will be regarded as a good one reviewers If these reviews are posted in a short period with consistent ratings... reviews stores Traditional methods what can you conclude from these features? review features rating: 4, bag-of-words: {switch, nook, kindle, why, what, learn} reviewer features name: KKX; number of reviews: 1 average rating: 4 store/product features Kindle average rating: 4/5 stars, price: $199 group spamming KKX wrote one review only, failed the frequency test