EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Self-Supervised Time-Aware Transformer for Credit Card Fraud Detection

Jiwon Kim, Hyunwoo Choi, Seojun Kang

International Journal of Business and Social Science Studies · 03.09.2026 · 6 (2) · 184-195

DOI: 10.66795/ijbsss71 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Business and Social Science Studies, online erschienen am 03.09.2026. Verfasst von Jiwon Kim und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Credit-card fraud detection is a temporally structured rare-event problem in which transaction labels are expensive and irregular inter-arrival times carry behavioral information. This paper presents a self-supervised time-aware Transformer (SSL-TAT) that couples a gated relative-time attention bias with two label-free pretext tasks: masked transaction reconstruction and local order-corruption detection. The temporal gate allows the encoder to learn when elapsed time should influence attention instead of imposing a fixed decay. We evaluate the method on a reproducible synthetic card-present transaction stream generated from fixed customer and terminal profiles with injected fraud episodes. The stream contains 81,558 transactions, including 1,349 fraud events, and is evaluated using chronological train/validation/test partitions with repeated random seeds. In the full-label sequence experiment, the time-aware Transformer improves the vanilla Transformer from 0.9799 to 0.9815 ROC-AUC and from 0.8900 to 0.9006 F1. SSL-TAT reaches 0.9813 ROC-AUC, 0.9600 AUC-PR, and 0.9083 recall, the highest recall among the evaluated neural sequence models. Tree ensembles remain strongest on this feature-rich synthetic benchmark, highlighting that the principal value of the proposed design lies in temporal representation learning rather than replacing competitive tabular learners. These results provide a reproducible basis for studying self-supervision, irregular timing, and sequence modeling in fraud detection.

Zitieren

APA

Kim, J., Choi, H., & Kang, S. (2026). A Self-Supervised Time-Aware Transformer for Credit Card Fraud Detection. International Journal of Business and Social Science Studies, 6(2), 184-195. https://doi.org/10.66795/ijbsss71

Vancouver

Kim J, Choi H, Kang S. A Self-Supervised Time-Aware Transformer for Credit Card Fraud Detection. International Journal of Business and Social Science Studies. 2026;6(2):184-195. doi:10.66795/ijbsss71

BibTeX
@article{Kim2026,
  author = {Kim, Jiwon and Choi, Hyunwoo and Kang, Seojun},
  title = {A Self-Supervised Time-Aware Transformer for Credit Card Fraud Detection},
  journal = {International Journal of Business and Social Science Studies},
  year = {2026},
  volume = {6},
  number = {2},
  pages = {184--195},
  doi = {10.66795/ijbsss71}
}
Kostenlos starten →