Budgeted Class-Balanced Drift Gating for Directional Temporal Graph Fraud Detection

Concept drift can render a graph-based fraud classifier obsolete even when its topology encoder remains useful. This study investigates when a delayed-label alarm should update a classifier under independent class-prior fluctuations and abrupt changes in transaction-flow roles. Building on the seven-channel directed two-hop simplified graph convolution and stable–plastic classifier of ATGNN-SP, we introduce BCG-5: a class-balanced predictive-loss gate, a five-update stream budget, and recent-only plastic fitting. Controlled experiments use reproducibly generated directed transaction graphs, 49 chronological snapshots, and five paired seeds in three regimes; no production transaction data are used. In the mixed regime, BCG-5 achieves average precision 0.3175 ± 0.0366 and F1 0.3413 ± 0.0693, compared with 0.3004 ± 0.0280 and 0.3367 ± 0.0712 for the loss-gated ATGNN-SP update rule, while reducing mean updates from 8.0 to 5.0. A plastic-only alternative achieves higher predictive accuracy at greater update frequency, revealing a meaningful stability–adaptability trade-off.