DriftGuard-HCL: Concept-Drift-Aware Continual Heterogeneous Graph Contrastive Learning for Evolving Financial Fraud Detection

Abstract

Financial fraud detectors are commonly trained as stationary classifiers even though transaction distributions, merchant populations, and attack typologies evolve. This paper presents DriftGuard-HCL, a continual heterogeneous graph-contrastive framework that combines a scalable typed graph-context encoder, semantic-group contrastive regularization, delayed-label replay, and a relation-support-aware drift gate. Each transaction is modeled on a temporal customer–category–merchant graph; causal customer and merchant neighborhood summaries are fused with a typed relation embedding through a bilinear interaction. A multi-signal detector combines Jensen–Shannon divergence over category and amount distributions, latent feature shift, and novel-relation mass. The detector switches a stability–plasticity ensemble from a frozen warm-up anchor toward a replay-regularized current model only after a calibrated shift. We evaluate the method with a strict prequential protocol on the public BankSim benchmark: 30 six-day snapshots, a one-snapshot label delay, five shared random seeds, and no future-label feature construction. In the native chronological stream, DriftGuard-HCL obtains 0.732 PR-AUC and 0.648 F1. In a controlled emergent-typology stress stream, it obtains 0.747 PR-AUC and 0.668 F1, improving over a continually updated heterogeneous contrastive model by 0.169 PR-AUC and 0.306 F1. The drift detector triggers once at typology restoration and never in the native stream. These results are reproducible from the accompanying code and raw result files; they support the value of drift-gated adaptation while not constituting evidence of production-bank performance.