Contrastive graph neural network-based camouflaged fraud detector

Recently GNN-based graph machine learning methods have been widely used in fraud detection. Prior works have revealed the problems of camouflaged fraudsters on graphs, which dilute the suspicion of aggregated features of fraudsters. The fraudsters’ camouflage lies in two main aspects: 1) fraudsters adjust their attributes to camouflage original hand-craft features against detection. 2) fraudsters connect to benign entities, making the labels of their contexts inconsistent. However, previous neighbor sampling methods based on similarity measures are inadequate in calculating neighbor similarity, resulting in a loss of GNN-based detector performance. In this paper, we enhance the node similarity measure-based neighbor sampling method and the multi-relational aggregator. First, we use supervised contrastive learning to aggregate discriminative node embeddings. After that, the consistent neighbors are sampled according to the contrastive embedding similarity as well as the node feature similarity. Finally, we improve the multi-relational aggregator via contrastive embeddings. Experiments on two real-world datasets demonstrate that our model CACO-GNN achieves the highest scores in F1-macro, AUC, and Recall, with an 18.5% improvement in F1-macro performance, enabling more effective detection of camouflaged fraudsters.