Collusion-aware detection of review spammers in location based social networks

Abstract

To ensure the quality of online review, more and more location-based social networks (LBSNs), like Yelp, have established the filtering systems to detect groups of review spammers. This is not an easy task. Review spammers use camouflage methods to maintain their spam behavior in a very low density to try to conceal themselves in normal users. These camouflaged spammers, driven by profits, are hired by some stores to write fake reviews in groups so as to raise these stores or to belittle their competitors. To avoid the unhealthy competition, in this paper, we propose a novel detection mechanism to discern collusive review spammers, including individuals and groups. The key point of our mechanism is to identify hidden spammers through multiple anomalous relationship features, especially the collusive relation between review spammers and the business competition between locations. Based on multi-view anomalous features, two detection models are proposed for individual and group discovery, respectively. For malicious individuals, a detection model based on Markov Random Field (MRF) is constructed to formalize an inference problem, where the corresponding marginal distribution of users and locations are calculated respectively. For review spammer groups, a hierarchical agglomerative clustering algorithm is conceived according to a new validity index to make sure the collusion relation in each group is close at most. Experiment results show that our method can detect collusive spammers and groups more accurately and comprehensively over the current researches. The additional experiments also show the effectiveness of each anomalous feature in detecting review spammers.