Continuous-Time Heterogeneous Graph Contrastive Learning for Financial Fraud Detection

Abstract

Financial fraud unfolds as a stream of typed interactions whose timing, relation semantics, and structural context jointly determine risk. Snapshot-based heterogeneous graph models preserve relation types but quantize time, whereas continuous-time temporal graph models usually emphasize event order without explicitly separating relation-specific behavioral states. This paper proposes CT-HGCL, a continuous-time heterogeneous graph contrastive learning framework centered on an event-driven multi-timescale state bank. For each node and relation, CT-HGCL maintains exponentially decayed count and amount states at fast and slow time scales; the state is updated only when an event arrives, so no temporal snapshots are required. A type-aware event encoder combines these causal states with endpoint types, relation identity, amount, absolute time, and inter-event gaps. Self-supervised pre-training forms two semantics-preserving stochastic views of each event state and optimizes an InfoNCE objective before weighted supervised fine-tuning. The design directly follows the continuous-time direction motivated by recent dynamic heterogeneous fraud research while replacing snapshot recurrence with exact-time state evolution. On a controlled event-stream benchmark containing 18,480 typed transactions, 4,195 active nodes, four relation types, and 480 fraud events organized into 120 collusive rings, CT-HGCL achieves 0.9596±0.0045 AUC-ROC over three random seeds. More importantly, at 5% labeled training data it reaches 0.8963±0.0253 AUC, compared with 0.8785±0.0576 for the same continuous-time heterogeneous encoder trained from scratch. Removing relation/type information reduces AUC to 0.9375±0.0046, showing that temporal continuity alone is insufficient. These results support continuous-time, heterogeneity-aware self-supervision as a practical route for label-scarce fraud detection.