Dual Channel Graph Neural Network for Fraud Detection

Fraud detection based on graph neural networks has received wide attention in recent years. In fraud detection, there exists a class imbalance problem because the number of fraudulent nodes is often much smaller than the number of normal nodes. Besides, fraudulent nodes usually camouflage themselves feature-wise and structure-wise, making the fraudulent information hidden from the normal information. This also introduces inconsistency to graphs. There must be edges connecting nodes with different labels, which violates the homophily assumption of vanilla GNNs. To address the above problems, we propose a Dual-Channel Graph Neural Network (DCGNN) for fraud detection. Firstly, we use the Class-balanced Node Sample Module for sampling so that the model can better learn the patterns of minority class nodes. Then, we design the Attribute-Structure Dual-Channel Hybrid Module so that the model adaptively combines the representations of two channels for detection. A graph disparity convolutional network is also introduced to model the dissimilarity between nodes to solve the inconsistency problem. Experiment results on the benchmark datasets demonstrate the effectiveness of our model.