Deceptive review detection using GAN enhanced by GPT structure and score of reviews

Abstract

These days, in online E-commerce platforms, the most of users utilize the purchase experience of other users reported in the reviews to make a right decision. As much as the positive effect of this phenomena, there would be a motivation to produce deceptive reviews toward some malicious goals. Therefore, the detection and deletion of these deceptive reviews from online platforms will be an essential task to keep them safe. Different approaches from Machine Learning (ML)-based methods to the recent Neural Networks (NN)-based ones have been proposed to detect deceptive reviews. Here, the lack of sufficient labeled data is an essential barrier, to obviate that Generative Adversarial Networks (GAN) have been utilized to generate some data with a distribution close to original ones. In this regard and along with the successful results of Generative Pre-trained Transformers (GPT) in textual tasks, it have also been used besides GAN framework to detect deceptive reviews. Despite a lot of efforts on this matter, there is no efficient study for considering metadata or behavioral features besides powerful generative models. In this paper, we have proposed a new approach called Score_Gpt2ganto consider the scores of reviews as a regularization concept besides the GPT-based GAN approach. Evaluation results in comparison between different methods have shown an increase in the accuracy of 1.4% on the TripAdvisor dataset and 3.8% on the YelpZip dataset by our proposed method.