Opinion Spam Detection through User Behavioral Graph Partitioning Approach

Abstract

Online reviews, an important source of user opinions, help not only other customers to make a decision but also manufacturers to improve quality of their products or services. Due to commercial reasons, untruthful reviews (spam) written to promote or demote certain products rather than they deserve have become a crucial problem. Although existing supervised approaches have shown the effectiveness of spam detection by using statistical learning, they require much expensive cost for labeling the training data. In this paper, we present BeGP that is a graph-partitioned approach for opinion spam detection. A set of characteristic features is first extracted and a user behavioral graph is constructed by connecting reviewers sharing those features to capture their similar behavior. BeGP is a semi-supervised scheme without requiring any training. Hence, it starts with a small subgraph of labeled spammers and afterwards iteratively expands by conducting connected other users as a resulted set of suspects. We demonstrate the effectiveness of BeGP on two real-world review datasets from Yelp.com. The result shows that it outperforms several state-of-the-art methods with accurately identifying spammers as well as review spams within the k-first order of ranking.