Identifying review spam with an unsupervised approach based on topic abuse

Abstract

The harmfulness of review spam (also known as deceptive opinion) has long been recognized. However, due to the lack of supervised annotations, detecting these fake reviews is challenging ever since the dawn of this field. In this paper, by exploring distinct composing patterns between sincere reviewers and spammers, we propose a novel approach to examine review contents and hunt long spams. Correlation levels upon product metadata and nominated aspects are highlighted for feature selection. We take two highly acknowledged metrics, i.e., duplication and burstiness, to evaluate our approach. Comparative results upon the top two Chinese business-to-customer websites show that our approach is effective and outperforms state-of-the-art solutions.