Graph transformer-based self-adaptive malicious relation filtering for fraudulent comments detection in social network

In recent years, Graph Neural Networks (GNNs) have proven to be effective in detecting fraud within social networks by gathering information from neighboring nodes to predict fraudulent actions. However, the continuous evolution and camouflaging tactics of fraudsters, such as establishing malicious connections with legitimate users and mimicking their behaviors, can often elude GNN-based detection methods. To tackle this issue, we propose a multi-order moments graph transformer, named MMGT, to effectively learn node representations through an attention mechanism that captures information of different orders of moments. Building upon this foundation, we further introduce a self-adaptive malicious relation filtering model for fraud detection, denoted as SFGT. Initially, we form a multi-relational graph that encapsulates complex relations within a given social network and derive node representations by accumulating neighbor node and edge data based on an enhanced graph transformer. Subsequently, a threshold-based malicious relation filter mechanism is proposed to eliminate malicious links by assessing the distance between nodes. Furthermore, an adaptive threshold learning policy is developed to bolster the model's performance and its ability to generalize. Finally, extensive experiments are carried out on two public datasets, Amazon and YelpChi, which underscores the effectiveness of our proposed model. The experimental results indicate that our model achieves state-of-the-art performance.