Adaptive Temporal Graph Neural Networks for Real-Time Financial Fraud Detection under Concept Drift

Abstract

Financial fraud detectors operate in a non-stationary environment in which illicit actors adapt, class prevalence changes, and labels arrive after investigation. This paper proposes ATGNN-SP, a drift-gated stable-plastic temporal graph neural network for real-time-compatible fraud scoring. Each directed transaction snapshot is converted into a 133-dimensional representation by a two-hop simplified graph convolution that separately propagates incoming, outgoing, second-order, and cross-direction information. A frozen stable multilayer perceptron preserves the initial concept, while a plastic copy is updated only when delayed prequential log loss exceeds a validation-derived control limit. Recent snapshots provide rapid adaptation and an anchor replay buffer limits forgetting. Experiments are executed on the real Elliptic Bitcoin transaction graph using a strict chronological protocol: time steps 1-29 for initial training, 30-34 for validation, and 35-49 for predict-then-update testing. Across three seeds, ATGNN-SP reaches F1 0.2649+/-0.0092 and average precision 0.2624+/-0.0193, improving the static graph MLP by 4.27% and 5.43% relative, respectively. A graph-aware gradient-boosting baseline obtains a slightly higher F1 of 0.2714, which is reported explicitly. The accompanying artifact contains all code, logs, and unedited result files; no experimental metric is synthesized or manually altered.