EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Explainable Cloud-Native Payment Fraud Detection Using SHAP-Guided Feature Adaptation

Sanne de Vries, Felipe Rocha

Computer Life · 20.07.2026 · 14 (3) · 13-20

DOI: 10.54097/12v3ax90 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer Life, online erschienen am 20.07.2026. Verfasst von Sanne de Vries und 1 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Real-time payment fraud detection in cloud-native environments must reconcile three competing pressures: predictive accuracy, strict tail-latency budgets, and growing regulatory demands for explainability. Prior work has mapped the accuracy-latency-footprint frontier for CPU-friendly classifiers but has not treated explainability, nor the runtime cost of the feature pipeline itself, as first-class design concerns. We introduce SHAP-Guided Feature Adaptation (SGFA), a teacher-student method that (1) trains a random-forest teacher on all features, (2) computes exact TreeSHAP attributions on a validation window to obtain a faithful global feature ranking, (3) selects a compact feature subset through a transparent rule—the smallest number of features reaching ninety percent of the teacher’s best validation PR-AUC—and (4) retrains a lightweight student on that subset. On the public ULB credit-card dataset (284,807 transactions, 0.173% fraud) under a chronological split and a precision-floor operating point, SGFA shrinks the feature set from thirty to eight (a 73% reduction) while a logistic-regression student retains 96.5% of the full-feature PR-AUC at a p99 latency of 321 microseconds and a 1.3 KB memory footprint; fifteen features fully recover full-feature accuracy. We further show that the SHAP ranking is competitive with the strongest filter baseline while additionally yielding faithful local and global explanations at no extra modeling cost, that it passes an AOPC faithfulness test (a 2.4x improvement over random ablation), and that the stability of attributions across time windows provides a practical concept-drift signal. SGFA thus demonstrates that explainability, far from being a tax on accuracy or latency, can actively improve a cloud-native fraud-scoring pipeline.

Zitieren

APA

Vries, S. D., & Rocha, F. (2026). Explainable Cloud-Native Payment Fraud Detection Using SHAP-Guided Feature Adaptation. Computer Life, 14(3), 13-20. https://doi.org/10.54097/12v3ax90

Vancouver

Vries SD, Rocha F. Explainable Cloud-Native Payment Fraud Detection Using SHAP-Guided Feature Adaptation. Computer Life. 2026;14(3):13-20. doi:10.54097/12v3ax90

BibTeX
@article{Vries2026,
  author = {Vries, Sanne de and Rocha, Felipe},
  title = {Explainable Cloud-Native Payment Fraud Detection Using SHAP-Guided Feature Adaptation},
  journal = {Computer Life},
  year = {2026},
  volume = {14},
  number = {3},
  pages = {13--20},
  doi = {10.54097/12v3ax90}
}
Kostenlos starten →