{Bongwon Suh Insoo Lee Woochang Hyun}

Abstract

Graph-based fraud detection faces significant challenges, such as severe class imbalance, inconsistent connections due to the scarcity of fraudulent nodes, and the camouflage of these nodes appearing like benign nodes. Existing studies often adopt the approach of filtering similar nodes to strengthen the homophily assumption of graph neural networks. However, to effectively address these issues, it is important to distinguish and adaptively utilize the labels of neighboring nodes. In this study, we propose the Label-Exploring Graph Neural Network (LEX-GNN), designed to enhance fraud detection by actively leveraging labeled node information. The core idea is that the manner of message passing and reception should vary depending on the node types. Specifically, we first predict the labels of nodes based on their original or previous representations. Subsequently, each node transmits differently processed messages according to its probability of being fraudulent. Finally, target nodes also receive the messages differently depending on their pre-predicted probability. Extensive experimental results on real-world benchmarks demonstrate that LEX-GNN outperforms existing state-of-the-art baselines. Our code is available at https://github.com/wdhyun/LEX-GNN.