Uncertainty-aware spatio-temporal contrastive graph neural networks for cyber financial fraud detection and risk management

Financial fraud detection requires screening massive transaction networks where evolving topologies, extreme label sparsity, and asymmetric misclassification costs make traditional classification paradigms ineffective. We propose ST-CGNN, a spatio-temporal contrastive graph neural network that frames operational screening as a multi-task learning problem in which a shared encoder is supervised by a contrastive regularizer and an evidential triage head. Concretely, ST-CGNN combines a continuous-time heterogeneous encoder with a hard-negative contrastive regularizer and an evidential output head, so that structural representations and uncertainty-aware prioritization are trained from a common backbone with summed losses rather than as a sequential, modular pipeline. Evaluated under strict chronological constraints on large-scale public and controlled benchmarks, ST-CGNN consistently outperforms state-of-the-art GNNs and post-hoc calibration methods. Specifically, on the DGraph-Fin benchmark, the proposed evidential triage score improves Precision@100 to 0.884 and achieves a calibrated ECE of 0.034. On Elliptic, the difference between ST-CGNN and the best competitor (MTP-GAT) lies within seed variance and is not statistically distinguishable; gains concentrate on benchmarks where heterogeneity and bursty timing dominate. Paired bootstrap tests and selective-prediction analysis confirm that this shared-encoder design significantly enhances the reliability of fixed-budget analyst reviews, providing a robust foundation for high-stakes risk management in dynamic transaction environments.