A Multi-Perspective Dual-Stream Framework Integrating Transformers and Graph Neural Networks for E-Commerce Opinion Fraud Detection

Abstract

Online word-of-mouth significantly shapes consumer purchasing decisions in the digital marketplace. However, malicious actors exploit this trust by deploying coordinated fake review campaigns that artificially inflate or deflate product popularity. Traditional detection approaches primarily analyze textual content in isolation, failing to capture the complex behavioral and relational patterns characteristic of organized fraud. This paper introduces H-TGNN (Hybrid Transformer-Heterogeneous Graph Neural Network), a novel multi-perspective dual-stream framework that simultaneously models linguistic content, user behavior, and structural relationships within e-commerce ecosystems. The architecture integrates transformer-based semantic encoders with heterogeneous graph attention networks across user-review-product topologies, enabling the identification of sophisticated, coordinated review farms. A comprehensive evaluation across three diverse benchmark datasets demonstrates superior performance, achieving a peak accuracy of 93.9% and an F_1-score of 93.5%, significantly outperforming contemporary state-of-the-art baselines. Statistical significance testing (p < 0.001) confirms the robustness of these architectural improvements. Furthermore, the framework demonstrates strong cross-platform generalization maintaining an 88%–91% accuracy window on entirely unseen platforms and exhibits linear scalability. Finally, a quantitative interpretability assessment backed by SHAP analysis and a formal user study with e-commerce platform moderators yields an 88% validation accuracy and a 4.2/5.0 explanation usefulness rating. The HTGNN framework advances the state of e-commerce fraud detection by synthesizing deep semantic analysis with structural network intelligence, ensuring digital platform integrity and consumer safety.