Graph for fraud detection

This post details the application of Relational Graph Convolutional Networks (RGCN) for fraud detection in Grab's ecosystem. It explains why graph-based models are superior to rule engines and decision trees for unknown fraud detection due to their ability to utilize structural correlations and require less feature engineering. The post presents the RGCN architecture, its semi-supervised learning approach, and performance metrics (ROC curve, AUROC). It also provides practical tips for using RGCN, such as limiting convolutional layers and utilizing rich node features. Furthermore, it discusses graph explainability through visualization and highlights ongoing challenges and future work, including feature initialization, real-time prediction, and handling noisy connections.