SSD-GNN: Fraud Detection based on Spectral and Spatial Dual Graph Neural Networks

Graph Neural Networks (GNNs) have attracted much attention due to their outstanding performance in fraud detection tasks, which reveal fraudsters by aggregating neighborhood features. However, most GNNs assume that neighborhood features are homogeneous, which does not match the fundamentally heterogeneous fraud graph, while heterogeneous information is equally important for fraud detection. In addition, the label-imbalanced neighborhood also deteriorates fraud detection accuracy. To address the above issues, we propose an innovative spectral and spatial dual graph neural network for fraud detection, namely SSD-GNN. In SSD-GNN, we design a dual-pathway framework, including spectral and spatial processing paths, and aggregate neighborhood features according to the homogeneity and heterogeneity assumptions, respectively. In the spectral processing path, we design high-pass filtering and low-pass filtering to form a mixed-pass filter to extract low-frequency information of homogeneity and high-frequency information of heterogeneity. In the spatial processing path, we first enhance the connection between the target node and the global node by clustering information to alleviate the label imbalance problem. Then we design similarity aggregation and disparity aggregation to capture homogeneity and heterogeneity information, respectively. We run experiments on two real-world fraud detection datasets, using AUC and F1-macro as test metrics. The comprehensive experimental results show that the AUC and F1-macro on the average source exceed at least 5 percent. demonstrate the superiority of the proposed SSD-GNN.