Answers are grounded in this paper and cite their sources.
Research on credit card fraud detection has rapidly increased over the past decade due to the increasing use of digital payments and the financial losses caused by fraud. Despite extensive research, most studies focus on detection performance, while transparency, fairness, and auditability remain under-studied, although they matter under regimes such as the General Data Protection Regulation (GDPR) and, for credit-scoring uses of fraud scores, the EU AI Act. This review argues that accountability, the ability to explain, audit, and demonstrate fairness in automated decisions, is under-explored compared to detection performance, and performance-related limits (imbalance handling, generalization, and offline-only testing) also remain unaddressed. To address this gap, a systematic literature review is conducted based on the PRISMA 2020 statement and includes papers published from 2016 to 2026 retrieved from IEEE Xplore, Scopus, ACM Digital Library, Web of Science, and Google Scholar. Five review questions (Section 2.1) framed the study and data extraction. A total of 45 papers were retained from 115 records (110 database records plus 5 citation-search records) after full-text assessment of 50 reports; 39 are primary studies (peer-reviewed publications and preprints), and 6 are surveys or fairness- methodology papers kept as background only. The primary studies were then divided into three areas: traditional machine learning (16), deep learning and Transformer-based models (15), and explainable AI and fairness (8). The synthesis shows consensus that ensemble and deep methods achieve high detection accuracy under class imbalance, but findings conflict on which imbalance handling strategy is most effective and on whether reported gains generalize beyond a small set of benchmark datasets. Research on explainability using techniques such as SHAP and LIME has emerged, yet interpretability remains appended to models rather than embedded in them, and formal fairness analysis remains limited across the three themes. Building on the findings, this review introduces EFMDS, a conceptual framework that has not yet been empirically implemented or validated, which is designed as a five-layer pipeline that integrates adaptive class-imbalance handling with multi-model detection, multi-method explainability, and formal fairness auditing, including closed-loop bias correction. This framework is intended to guide the development of a fraud detection system that is robust, interpretable, and equitable.
No takes yet. Share an insight, caveat, or question.
Mahlake et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: