Dynamic heterogeneous graph contrastive learning for uncovering collusive financial fraud

Detecting collusion rings in modern banking requires modeling the evolving structural interactions among heterogeneous entities (customers, accounts, and devices) rather than isolated transaction features. Most graph-based fraud detectors assume abundant labels, yet confirmed fraud labels in anti-money laundering (AML) settings routinely arrive months after the fact. We introduce Audit-HCL, a dynamic heterogeneous graph neural network framework that uses dual-view contrastive learning to operate effectively under this label scarcity. Audit-HCL represents the transaction ecosystem as a temporal sequence of heterogeneous graph snapshots, encodes them through a metapath-guided heterogeneous attention encoder, and tracks evolving node behavior with a GRU-based temporal dynamics module. A cross-view contrastive objective aligns structural and temporal perspectives for legitimate nodes while separating anomalous ones, guided by an anomaly-aware negative sampling strategy. Experiments on two public benchmarks (Elliptic and IBM AML-Synthetic) show that Audit-HCL outperforms fourteen baselines by 3.2% in AUC-ROC and 6.8% in F1-score, with the gains over the strongest competitors confirmed by paired significance tests, and that it retains useful discriminative power with zero fraud labels. On the synthetic IBM AML benchmark, it also detects laundering patterns an average of 7.4 weeks ahead of confirmed events ([Formula: see text] the lead time of the best baseline) by capturing gradual structural drift before large-scale fund transfers begin, although the magnitude of this lead time is tied to the controlled typologies of the synthetic data and should be read as indicative rather than as a guarantee for production AML environments.