Metapath and syntax-aware heterogeneous subgraph neural networks for spam review detection

Spam Review Detection is a subclass of text classification that aims to distinguish genuine reviews from spam reviews (e.g., irrelevant reviews, deceptive reviews, machine-generated reviews, and non-review messages). Previous studies have focused on review text analysis, abnormal behavior detection, and intrinsic relationship identification. However, these methods ignore different fraudulent camouflages and writing styles. In this paper, we instead design a Spam detection model Metapath-based Subgraph Aggregated Neural Network (Spam-MSANN) integrating three metapath-based subgraphs (i.e. User-Item Subgraph, Review Subgraph, and User-Review-Item Subgraph) and syntactic information to enhance the relevant representation of the review information with subgraph aggregation operations. Experimental results on benchmarking datasets (i.e. YelpCHI and Amazon) demonstrate that our Spam-MSANN model significantly improves the state-of-the-art models. Specifically, Spam-MSANN outperforms 11 of the 12 advanced benchmark models on these two datasets, which further manifests that the fusion of different metapath-based subgraphs and syntactic information is adequate for the spam review detection task.