Large-Scale Graph Neural Networks for Fraud Detection in Financial Transaction Networks: Architecture Design, Sampling Strategies, and Real-Time Inference at Billion-Edge Scale
Financial transaction networks -- in which nodes represent accounts and edges represent transactions -- exhibit graph-structural fraud patterns (ring networks, layering chains, velocity clustering) that are invisible to transaction-level classifiers but detectable through graph neural network architectures that aggregate multi-hop neighborhood information. This paper presents FraudGNN, a production-oriented Graph Neural Network system for real-time fraud detection at billion-edge transaction graph scale, and reports its design, training, and deployment experience at a large payments processor. FraudGNN employs GraphSAGE with neighbor sampling for scalable inductive inference, incorporating temporal edge features, account behavioral embeddings, and network centrality features into a heterogeneous graph transformer architecture. Key engineering contributions include a mini-batch training pipeline supporting 8-billion-edge graphs on 64-GPU clusters using gradient checkpointing and heterogeneous graph partitioning, and a real-time inference serving architecture that delivers GNN predictions within 45ms P99 latency for payment authorization decisions. FraudGNN achieves 94.3% AUC on a held-out fraud detection benchmark, representing an 8.7 percentage point improvement over the XGBoost baseline. We characterize the graph data pipeline engineering challenges -- including temporal graph construction, feature freshness management, and graph store selection -- that represent the majority of production deployment effort. This paper provides the most complete engineering treatment of production-scale GNN fraud detection systems published to date.
