Knowledge-Guided Fraud Detection Using Semi-supervised Graph Neural Network

Abstract

Fraud detection is about finding unusual behaviors in the data, and it is essential for companies to detect fraudulent users to prevent unpredictable risks. The graph-based approaches model relationships into graphs for capturing the intricate characteristics of complex scenarios to detect fraudsters. However, it still faces the problem of data skew where labeled fraudsters are far fewer than unlabeled examples. Knowledge may help identify these unlabeled data, thus this paper combines domain knowledge with GNN and proposes a Knowledge-Guided Semi-supervised Graph Neural Network, namely KS-GNN, to address the problem of data skew. We utilize domain experts to design small amount of rules to roughly label unlabeled data as noisy and use a semi supervised method to train fraud detectors. By utilizing only 13 GFD rules conducted by domain experts, the performance of our method yields about 15% improvement over the state-of-the-art fraud detection methods CARE-GNN on banking transaction funds supervision datasets (BTFSD). Moreover, with some modification of the GFD rules on BTFSD, the performance of KS-GNN on other domain datasets such as IEEE-CIS Fraud Detection (https://www.kaggle.com/c/ieee-fraud-detection/data) and Yelp-Chi is also improved by about 5% on average compared with the state-of-the-art methods.