GraphTFD: A Fraud Detection System Based on Graph Transformer

Chengcheng Yu, Lujing Fei, Wenqian Zhou, Lin Chen, Jiahui Wang, Xiankai Meng, Fangshu Chen

Abstract: Fraud detection is critical in identifying and preventing fraudulent activities across various industries, safeguarding financial assets, and maintaining trust in business. Although graph-based fraud detection methods have demonstrated significant success, they still suffer from two main challenges that remain unresolved. First, traditional GNN-based algorithms may perform poorly when the nodes’ class distribution is heavily skewed. Second, many GNN-based methods fail to generalize to the heterophily setting. In this demonstration, we present a graph transformer-based fraud detection system called GraphTFD. This system can help address the above challenges by exploring a data augmentation strategy to enhance the minority class with numerous unlabeled nodes and introducing group-level aggregation and group encodings to help the transformer encoder capture more similar neighborhood information under heterophily settings. GraphTFD also provides an interface that visualizes the GraphTFD internals and assesses the quality of fraud detection results. Our demonstration video can be found here: https://b23.tv/Z13L5Uf.

External IDs:dblp:conf/wise/YuFZCWMC24