Collusive Opinion Fraud Detection in Online Reviews: A Probabilistic Modeling Approach

摘要

We address the collusive opinion fraud problem in online review portals, where groups of people work together to deliver deceptive reviews formanipulating the reputations of targeted items. Such collusive fraud is considered much harder to defend against, since the participants (or colluders) can evade detection by shaping their behaviors collectively so as not to appear suspicious. To alleviate this problem, countermeasures have been proposed that leverage the collective behaviors of colluders. The motivation stems from the observation that colluders typically act in a very synchronized way, as they are instructed by the same campaigns with common items to target and schedules to follow. However, the collective behaviors examined in existing solutions focus mostly on the external appearance of fraud campaigns, such as the campaign size and the size of the targeted item set. These signals may become ineffective once colluders have changed their behaviors collectively. Moreover, the detection algorithms used in existing approaches are designed to only make collusion inference on the input data; predictive models that can be deployed for detecting emerging fraud cannot be learned from the data. In this article, to complement existing studies on collusive opinion fraud characterization and detection, we explore more subtle behavioral trails in collusive fraud practice. In particular, a suite of homogeneity-based measures are proposed to capture the interrelationships among colluders within campaigns. Moreover, a novel statistical model is proposed to further characterize, recognize, and predict collusive fraud in online reviews. The proposed model is fully unsupervised and highly flexible to incorporate effective measures available for better modeling and prediction. Through experiments on two real-world datasets, we show that our method outperforms the state of the art in both characterization and detection abilities.