Temporal Graph Learning for Fraud Detection in Large-Scale Financial Transactions

Abstract

Financial fraud detection remains challenging due to the dynamic nature of transaction behaviors and the complex relationships among financial entities. This study proposes a temporal graph learning framework that integrates transaction-level features with evolving relational structures to identify fraudulent activities in large-scale financial networks. The proposed approach combines temporal attention mechanisms with graph neural networks to capture both short-term behavioral changes and long-range transaction dependencies. Experiments are designed around a dataset containing approximately 1.2 million financial transactions collected over 18 months, including 38,400 labeled fraudulent activities. The proposed model achieves an illustrative AUROC of 0.963, a precision of 0.831, and a recall of 0.762, outperforming conventional machine learning baselines such as Random Forest and XGBoost. Compared with the strongest baseline, the framework improves precision-recall AUC from 0.722 to 0.791 while maintaining an average inference latency of 24 milliseconds per transaction. These findings suggest that temporal graph representations can improve fraud identification by capturing previously overlooked behavioral dependencies in financial transaction networks.