Loss-Responsive Dual-Memory Graph Models for Fraud Drift: A Controlled Temporal Stress Test (Journal of Computing and Electronic Information Management)

Summary

• The reproducible synthetic temporal transaction stream generator fixes its random seed to 20261022 and creates T=49 directed snapshots with 240 nodes each, totaling 11,760 node instances.p.2
• The adaptive mixture triggers updates on all 15 test snapshots, compared with 7.8 updates for the fixed mixture baseline.p.1

Key insights

1. 1. Exponential expert weighting in the adaptive dual-memory mixture improves mean average precision to 0.4170 from 0.3469 in fixed dual-memory fusion.p.6
2. 2. Retaining a frozen stable predictor in the final score does not inherently improve outcomes compared to plastic-only adaptation.p.1
3. 3. Scoring predictions prior to model updating ensures an auditable chronological sequence between predictive scores and delayed feedback.p.1
4. 5. A mismatch exists between the blended loss triggering retraining gates and the true prospective classification benefit of updating plastic weights.p.4
5. 6. Expert mixture weight recalculation requires only two exponentials and constant memory, adding negligible computational overhead relative to graph backpropagation.p.4
6. 7. Direction-separated graph feature propagation highlights asymmetric transaction flows without conflating incoming and outgoing structural signals.p.6