Leveraging Multi-View Reviewer Representation for Spammer Detection in Online Reviews
Primary Investigator:
Julia (Taylor) Rayz
Shih-Feng Yang, Julia Taylor Rayz
Abstract
This poster presents a proposal and preliminary results by applying multi-view network representation learning for spammer detection on online review datasets collected from Amazon and Yelp. Previous research has focused on the textual characteristics of reviews, spammer behaviors, and the common patterns in a spammer group. A recent trend of utilizing network inference on spam detection enables us to jointly consider spam reviews, spammers, and spammer groups as heterogenous networks. However, several gaps exist in the literature on whether reviewer networks were properly designed for spam detection. First, previous studies considered only first-order proximity but did not include second-order proximity (Tang et al., 2015) when inferencing the structure of reviewer networks. Second, in real world, reviewers can form various network views with respect to different features. Previous studies treated reviewer networks as single-view networks rather than multi-view networks. In this study, we adopt MVE (Qu et al., 2017) to learn robust reviewer embedding from multi-view reviewer network. We use the learned reviewer embedding as reviewer’s features and investigate its effectiveness on spammer classification. Moreover, we conduct a cluster analysis on reviewer embeddings to evaluate the correlation between a reviewer’s spam tendency and its cluster.