An unsupervised domain adaptation method for cross-domain deceptive reviews detection

Abstract

Deceptive reviews seriously affect the interests of consumers, honest sellers, and e-commerce platforms. As e-commerce platforms often involve multiple domains (i.e., different products or services), in-domain deceptive review detection models trained and tested on a specific dataset may not perform well on other domains. Moreover, obtaining annotated data for so many individual domains is unrealistic. Cross-domain deceptive review detection aims to leverage labeled source domain data to improve the model’s performance on unlabeled target domain data. However, existing cross-domain deceptive review detection methods require labels for target domain data or do not consider domain-specific information. To further advance research, this paper proposes an unsupervised domain adaptation method for detecting cross-domain deceptive reviews. First, we propose a multiple mask views generation method to enhance domain-specific information to obtain different mask views of reviews. Secondly, the BERT and mask attention mechanisms are used sequentially to obtain contextual representations of the mask views and the original view of reviews. Thirdly, to maintain the consistency between the mask views and the original view of reviews, we use the intra-domain Kullback-Leibler divergence to constrain their learning process. Moreover, we use inter-domain dynamic maximum mean discrepancy and conditional maximum mean discrepancy to reduce differences between the distribution of source and target domains. Three sets of experiments on two datasets show that our method is superior to the baselines. In particular, the impact of domain differences on domain adaptability is further analyzed according to the quantified metric named domain distance defined in this paper.