Detecting collusive fake reviews using heterogeneous graph attention networks

Given that heterogeneous information networks contain richer information and more complex semantics compared to homogeneous graphs, this paper utilizes heterogeneous graphs to depict the diverse relationships among users, reviews, and stores within a fake review dataset, thereby more effectively revealing the associations between review publishers and their reviews. This structural advantage facilitates the detection of collusive behaviors among fake review groups and captures global features. Moreover, we employ a Heterogeneous graph Attention Network (HAN) for automatic feature extraction of reviewer nodes. Within this framework, node-level attention learns the interactions between nodes and their neighbors defined by meta-paths, while semantic-level attention focuses on assessing the importance of different metapaths in the heterogeneous graph for specific tasks. Through the learning of these two levels of attention, our model can hierarchically optimize the combination of neighbors and meta-paths, resulting in node embeddings that more accurately capture the complex structures and rich semantics within the heterogeneous graph. Based on these advanced features, our approach effectively detects and identifies collusive fake reviews, demonstrating superior performance.