Mitigating Human and Computer Opinion Fraud via Contrastive Learning

These platforms collect data about both users' and items' attributes, as well as accumulate the ratings and feedback of products and services, to develop algorithms for significant enhancement of users' experience on the marketplace. These algorithms are capable of influencing the purchasing behavior of users by (1) offering them the selection of the most relevant personalized positions, (2) reducing the individual searching costs, and (3) alleviating the information asymmetry on large commercial platforms with homogeneous sellers and products through feedback mechanisms. Since recommender systems have the power to affect the marketing decisions of users, they have become an attractive target for ratings and reviews manipulations, also known as attacks. Specifically, these attacks are aimed at inflating/deflating the ranks and text reviews of certain product positions or at simply sabotaging the efficiency and credibility of the the commercial platform in general. The current study focuses on solving the task of filtering out the deceptive opinions and detecting anomalous behavior on a platform with text reviews. The emphasis on text reviews can be explained by the fact that texts are a more informative and a more reliable source of product's and seller's quality, than a star-rating system, which is easy to manipulate (see [19], [14], [27], [28]).