Enhancing Fraud Detection via GNNs with Synthetic Fraud Node Generation and Integrated Structural Features

Abstract

Graph Neural Networks are widely employed for node classification in attributed networks. When it comes to fraud detection, however, GNNs can perform poorly, because a node’s features are typically computed based on its local neighborhood, and this allows fraudsters to “blend in” among legitimate users. In this paper, GNNs and supervised contrastive learning are proposed for fraud detection on datasets where fraudsters may use intricate strategies to camouflage themselves within the network. We train our GNNs using novel structural features in addition to those typically used in similar studies. The proposed features are based on the empirical probability distributions of various graph structural attributes which are extracted from a given dataset. We also apply supervised contrastive learning, enhanced with synthetic samples for the minority class (i.e., the fraudsters). Under our approach, the classifying capability of the GNN (measured via F1-macro, AUC, Recall) is improved by boosting the representation power of the calculated embeddings that maximize the similarity between legitimate users while minimizing that between fraudsters and legitimate users. Numerical experiments on two real-world multi-relation graph datasets (Amazon and YelpChi) demonstrate the effectiveness of the proposed method, whose improvements over the state-of the-art were especially significant in the larger YelpChi dataset.