DeepSpot: Understanding Online Opinion Spam by Text Augmentation using Sentiment Encoder-Decoder Networks

Abstract

Recently opinion spam has been widespread on online review websites and has received significant research attention. Existing approaches to detecting online opinion spam can be categorized into three groups: (1) review behavior-based approaches, which use metadata associated with user review behavior and product profile, (2) language-based approaches, which focus on the characteristics of the language that the opinion spammers use, and (3) graph-based approaches, where various user-review-product networks are constructed for node connectivity and similarity analysis. Unfortunately, all the aforementioned approaches have their limitations. In this paper, we introduce a holistic system, DeepSpot, for fake review detection. DeepSpot recognizes the true and fake reviews based on both the real human-posted reviews and the synthetic machine-generated reviews leveraging sentiment classification. Specifically, DeepSpot augments the original reviews with synthetic reviews using the encoder-decoder neural networks trained by the positive and negative reviews, respectively. Extensive experiments on real-world data showed that DeepSpot outperformed the state-of-the-art approaches in terms of various effectiveness metrics for recognizing true and fake reviews.