Combating Product Review Spam Campaigns via Multiple Heterogeneous Pairwise Features

Chang Xu, Jie Zhang

Abstract: Spam campaigns spotted in popular product review websites (e.g., amazon.com) have attracted mounting attention from both industry and academia, where a group of online posters are hired to collaboratively craft deceptive reviews for some target products. The goal is to manipulate perceived reputations of the targets for their best interests. Many efforts have been made to detect such colluders by extracting pointwise features from individual reviewers/reviewer-groups, however, pairwise features which can potentially capture the underlying correlations among colluders are either ignored or just explored insufficiently in the literature. We observed that pairwise features can be more robust to model the relationships among colluders since they, as the ingredients of spam campaigns, are correlated in nature. In this paper, we explore multiple heterogeneous pairwise features in virtue of some collusion signals found in reviewers' rating behaviors and linguistic patterns. In addition, an unsupervised and intuitive colluder detecting framework has been proposed which can benefit from these pairwise features. Extensive experiments on real dataset show the effectiveness of our method and satisfactory superiority over several competitors.