MSGFD: a multi-scale temporal–metric graph ODE network for financial fraud detection

Financial fraud detection on multivariate transaction streams is critical for safeguarding digital payments and credit risk management in modern financial platforms. Real-world financial fraud detection remains challenging due to severe class imbalance, evolving attacker behaviors, long and irregular transaction sequences, and complex interactions among high-dimensional financial indicators. Existing approaches, ranging from traditional sequence models to graph neural networks (GNNs) and transformer-based architectures, either compress transaction histories into static tabular summaries or focus solely on temporal dependencies, resulting in limited joint modeling of temporal evolution and metric-domain indicator relationships. Graph-based models typically construct entity-level transaction graphs rather than indicator-level relation graphs, making it difficult to capture multi-timescale numerical co-variation between indicators. Furthermore, conventional self-attention and deep GNN stacks often incur high computational and memory costs on long sequences with many features, hindering deployment under tight latency constraints. To address these issues, we propose MSGFD, a Multi-Scale Temporal-Metric Graph Ordinary Differential Equation (ODE) Network that jointly models temporal dynamics and value-driven indicator relations for financial fraud detection. MSGFD combines a linear-attention temporal encoder, a Multi-timescale Value-driven Relation Graph, a spatio-temporal graph ordinary differential equation metric encoder, and an attention-based sequence-level fraud head into a unified temporal–metric–temporal architecture. Experiments on two large real-world benchmarks, IEEE-CIS Fraud Detection and American Express Credit Default, demonstrate the effectiveness of MSGFD, improving macro F1-score and area under the precision–recall curve (AUC-PR) by up to 9.7% and 2.6% over the strongest baseline. Comprehensive ablation studies further verify the contribution of each component and the efficiency of the proposed framework.