EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Drift-Resilient Fraud Classification via Continual Learning with Replay-Guided Synthetic Minority Oversampling

Elena Rossi

Computer Life · 02.06.2026 · 14 (2) · 23-29

DOI: 10.54097/9t0znt37 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer Life, online erschienen am 02.06.2026. Verfasst von Elena Rossi. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Financial fraud detection systems operate under persistent conditions of distributional non-stationarity, where evolving fraud patterns and severe class imbalance jointly undermine the reliability of statically trained classifiers. This paper presents a drift-resilient fraud classification framework that integrates continual learning (CL) with a replay-guided synthetic minority oversampling technique (SMOTE) module to address both challenges through a unified training loop. The proposed system maintains an episodic memory buffer populated by strategically selected minority-class exemplars, which simultaneously serve as replay anchors for preventing catastrophic forgetting and as geometrically informative neighborhood seeds for cross-temporal synthetic sample generation. A drift-aware buffer management policy prioritizes boundary-proximal and recently misclassified minority instances, maximizing the informativeness of fixed memory capacity across successive temporal periods. Experimental evaluation on two benchmark fraud datasets under simulated concept drift conditions of varying magnitude demonstrates that the framework achieves a 12.4% improvement in area under the precision-recall curve (AUPRC) over standard gradient-boosted classifiers and a 9.7% gain over naive replay baselines. Ablation experiments confirm the independent contribution of each system component, with drift-aware buffer management identified as the single most impactful design choice. The results establish that tightly coupling memory-guided oversampling with experience replay constitutes a principled and effective strategy for robust, long-horizon fraud detection in non-stationary data streams.

Zitieren

APA

Rossi, E. (2026). Drift-Resilient Fraud Classification via Continual Learning with Replay-Guided Synthetic Minority Oversampling. Computer Life, 14(2), 23-29. https://doi.org/10.54097/9t0znt37

Vancouver

Rossi E. Drift-Resilient Fraud Classification via Continual Learning with Replay-Guided Synthetic Minority Oversampling. Computer Life. 2026;14(2):23-29. doi:10.54097/9t0znt37

BibTeX
@article{Rossi2026,
  author = {Rossi, Elena},
  title = {Drift-Resilient Fraud Classification via Continual Learning with Replay-Guided Synthetic Minority Oversampling},
  journal = {Computer Life},
  year = {2026},
  volume = {14},
  number = {2},
  pages = {23--29},
  doi = {10.54097/9t0znt37}
}
Kostenlos starten →