Review spam and reviewer behavior analysis

In recent years, opinion retrieval and mining attracted a lot of attention in the research community due to a wide range of applications. An important issue that has been largely ignored is the review spam, which refers to writing fake reviews to promote or to damage the reputations of some objects. Detecting such fake reviews and reviewers is critical for effective opinion mining and opinion search. In first part of this work we show different classification techniques used to detect various types of review spam with some degree of success. However, review spam appears to be harder to detect because it is very hard, if not impossible, to recognize fake reviews by manually reading them. So, in the second part of this work, we deal with a restricted problem, i.e. identifying unusual review patterns which can represent suspicious behaviors of reviewers, e.g., a reviewer who wrote all negative reviews on many products of a brand when other reviewers are generally positive about the brand is naturally suspicious. We believe that these patterns can provide useful signals for spam detection. We formulate the problem as finding unexpected rules and rule groups by borrowing some ideas from data mining. To find unexpected rules, one needs to know what is expected. A natural approach is proposed, which requires no user input. It defines expectations based on the inherent distribution of the data. Based on these expectations, metrics are designed to measure unexpectedness of rules and rule groups. The technique is domain and application independent. Using the technique, we analyzed an Amazon.com review dataset and found many unexpected rules and rule groups which indicate spam activities. To demonstrate the domain/application independent nature of the proposed method, we also analyzed a set of tweets from Twitter.com and found many pieces of interesting information.