A deep feature interaction and fusion model for fake review detection: Advocating heterogeneous graph convolutional network

In contemporary real-world scenarios, opinion spammers are hired to fabricate reviews that unfairly promote or demote particular products or services for personal gain. Although considerable attention has been devoted to addressing the problem, existing approaches often overlook the heterogeneous nature of reviewer–product interactions. Specifically, the correlation between review text (comments) and overall ratings, which provides various latent rich information to expose fake reviews, remains inadequately explored. Current methodologies focus on limited interactions, such as reviewer–review, product–review, or reviewer–product interactions, while neglecting significant aspects like reviewer–review–product and reviewer–rating–product interactions, leading to inadequate classifier performance. Motivated by this observation, this study proposes a novel Deep Feature Interaction and Fusion Model (DFIFM) whose ideas are five-folds: (a) constructing a reviewer–product interaction bipartite graph that represents heterogeneous feature node interactions through review text and overall rating values; (b) recognizing the existing mutual interactive relationship between review text and overall rating features, we construct a unified GCN to gain additional insights into feature relationships and capture mutual heterogeneous interactions between nodes; (c) to handle the encoding of unstructured review text features as edge attributes, we adopt a convolutional neural network (CNN); (d) attention mechanisms and fusion techniques are employed to capture interdependencies among reviewer–product latent features; and (e) a Multilayer Perceptron (MLP) utilizes the resulting latent feature representation for review classification. Experimental results on three publicly available datasets demonstrate its superiority over state-of-the-art baselines.