Graph Neural Networks for Real-Time Transaction Fraud Detection in Payment Networks
Modern payment fraud increasingly manifests through coordinated activity across cards, devices, merchants, customers, and IP addresses, making transaction-only models insufficient for detecting multi-entity attacks. This paper presents a real-time fraud detection framework based on a heterogeneous temporal graph neural network (HT-GNN) and a two-stage inference architecture that separates offline embedding computation from online transaction scoring. The model represents payment ecosystems as dynamic heterogeneous graphs, uses relation-specific attention and temporal aggregation to learn entity representations, and combines these embeddings with transaction attributes in a lightweight scoring head. On a largescale anonymized dataset containing 847 million transactions and 52 million entities, HT-GNN achieves 65.8 % recall at 0.1 % false positive rate, improving over strong non-graph and graph baselines while maintaining $\mathbf{4 3 ~ m s ~ P 9 5}$ latency. A full runtime analysis shows that the two-stage design constrains end-toend online scoring to 43 ms P95 and offline graph refresh to approximately 28 minutes on four V100 GPUs. Ablation studies show that heterogeneity, device/IP relationships, and temporal attention are the main drivers of performance gains. The results indicate that graph-based models are a practical direction for production-grade fraud detection in payment networks.