Cross-Domain Fake Review Detection via Orthogonal Counterfactual Representations

The popularity of online review-based purchases has led many businesses to generate fake reviews. Manual detection of these fakes is a challenging task. Moreover, existing research focuses primarily on single domain. Further, labelled datasets are limited and restricted to a few domains. Hence, detecting fake reviews in less-explored domains is a significant challenge. Cross-domain fake review detection can be used to handle these challenges by using the knowledge from labeled domains to improve detection in less explored domains. For this, a novel framework, COAT, that trains on Counterfactuals generated by the Orthogonal Alignment of BERT-driven mask infillings for cross-domain fake review detection has been proposed in the paper. Working of the proposed COAT is divided in three phases: (1) two-step masking to remove source-specific cues, (2) counterfactual generation using BERT, followed by orthogonal alignment via Procrustes algorithm, and (3) training and prediction using a Bi-LSTM classifier in a zero-shot setting. Experiments are conducted on benchmark datasets of Online Products, Hotels, Yelp Restaurants, Doctors and Restaurant reviews, divided in two groups. Obtained results show that the proposed COAT framework exhibits an enhanced domain transfer with a significant improvement in average accuracy by 17.31% and 26.18%, as well as F1-score by 19.05% and 51.16% on the two groups, respectively, over its counterparts in consideration. This shows that COAT empowers businesses, online platforms, and consumers with more reliable insights and fosters informed decision-making and trust in e-commerce platforms.