Know-GNN: An Explainable Knowledge-Guided Graph Neural Network for Fraud Detection

Abstract

Fraud is on the rise under modern e-commence scenarios, which will critically damage the market system. Thus, it is essential to detect fraudsters to prevent unpredictable risks. There are two challenges toward this problem. First, real world fraud detection usually lack of labeled samples. Second, recent ML-based detection method lack of interpretation. Knowledge may help with these problems. Hence, we propose a Knowledge-Guided Graph Neural Network, namely Know-GNN, which utilizes the expertise to roughly mark unlabeled data and uses an explainable semi-supervised method to train a fraud detector. We adopt Graph Functional Dependency (GFD) as a uniform expression of knowledge to mark unlabeled data and give explanations of the detection results. Experiments on banking transaction funds supervision data (BTFSD) demonstrate the effectiveness of our model. By utilizing only 13 GFD rules conducted by domain experts corresponding to BTFSD, the performance of our method yields about 14% improvement over the state-of-the-art methods, CARE-GNN. Moreover, the interpretable results can give interesting intuitions about the fraud detection tasks.