Temporal Graph Learning for Detecting Coordinated Financial Fraud in Transaction Networks

Abstract

Financial fraud increasingly involves coordinated activities distributed across multiple accounts, making isolated transaction analysis insufficient for identifying complex fraudulent behavior. This research investigates a temporal graph learning approach that models financial transactions as evolving networks of interacting entities. Rather than relying exclusively on transaction-level characteristics, the proposed method combines account connectivity, transaction timing, and behavioral irregularities to identify suspicious activity patterns. The evaluation uses a simulated financial network containing 2.8 million transactions, 185,000 accounts, and a fraud prevalence of approximately 0.42%. Comparative experiments examine gradient boosting, recurrent neural networks, and conventional graph neural networks. The proposed model achieves a precision-recall AUC of 0.512, outperforming the strongest baseline by 0.083. At a fixed investigation capacity covering the highest-risk 2% of transactions, the model identifies 68.7% of fraudulent events, compared with 55.4% using transaction-based classification. Additional analysis indicates that incorporating temporal relationships improves the identification of previously unseen fraud clusters. The results highlight the potential of dynamic relational modeling for financial risk surveillance.