Real-time dynamic graph learning with temporal attention for financial fraud detection

Financial transaction risk control is a cornerstone of intelligent finance platforms, yet existing approaches remain limited. Early frameworks modeled user behaviors independently, while later graph-based systems extracted handcrafted features from capital-flow networks. Although these methods improved detection, they struggle to capture fine-grained temporal dynamics and evolving topological patterns, and they depend heavily on manual feature engineering. In this work, we present a unified real-time dynamic graph learning framework that directly learns representations from raw streaming transaction graphs. Central to our design is a continuous-time, context-aware graph attention transformer (C2GAT), which models both higher-order structural dependencies and temporal patterns. We further decouple multi-role interaction paths and local neighborhood structures into dedicated subgraph modules, enabling complementary views of fraud behaviors. Evaluated on an industrial credit-cashback fraud detection scenario, our framework delivers substantial improvements in accuracy and false-alarm reduction over industry-standard baselines, while meeting stringent real-time latency requirements for deployment in large-scale financial systems.