Decoupling Graph Neural Network with Contrastive Learning for Fraud Detection

Recently, many fraud detection models introduced graph neural networks (GNNs) to improve the model performance. However, fraudsters often disguise themselves by camouflaging their features or relations. Due to the aggregation nature of GNNs, information from both input features and graph structure will be compressed for representation learning simultaneously. On the one hand, since not all neighbors provide useful information due to camouflage, aggregating information from all neighbors may potentially decrease the model performance. On the other hand, the structure including all neighbors is not reliable due to the relation camouflage. In this paper, we propose to decouple attribute learning and structure learning to avoid the mutual influence of feature and relation camouflage. Therefore, the model first learns its embedding seperately and then combine them together with label-guided contrastive losses to make predictions better. We conduct extensive experiments on two real-world datasets, and the results show the effectiveness of the proposed model.