A Deep Hybrid Model for fake review detection by jointly leveraging review text, overall ratings, and aspect ratings

Recently, product/ service reviews and online businesses have been similar to the blood–heart relationship as they greatly impact customers’ purchase decisions. There is an increasing incentive to manipulate reviews, mostly profit-motivated, as positive reviews imply high purchases and vice versa. Therefore, a suitable fake review detection approach is paramount in ensuring fair e-business competition and sustainability. Most existing methods mainly utilize discrete review features such as text similarity, rating deviation, review content, product information, the semantic meaning of reviews, and reviewer behaviors. In the matter of discourse, some recent researchers attempted multi-feature (review- and reviewer-centric features) integration. However, such approaches face two issues: (1) Review representation is extracted in an independent manner, thus ignoring correlations between them (2) Lack of a unified framework that can jointly learn latent text feature vectors, aspect ratings, and overall rating. To address the named issues, we propose a novel Deep Hybrid Model for fake review detection, which jointly learns from latent text feature vectors, aspect ratings, and overall ratings. Initially, it computes contextualized review text vectors, extracts aspects, and calculates respective rating values. Then, contextualized word vectors, overall ratings, and aspect ratings are concatenated. Finally, the model learns to classify reviews from such unified multi-dimensional feature representation. Extensive experiments on a publicly available dataset demonstrate that the proposed approach significantly outperforms state-of-the-art baseline approaches.