Abstract
Financial institutions struggle to achieve both effective fraud detection and equitable outcomes under new regulations like the EU AI Act. This study evaluates a comparative resampling-debiasing framework for credit card fraud systems using a 100,000-transaction dataset with extreme class imbalance (1:99). Through a comparative analysis of Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), and Random Oversampling (ROS) with Extreme Gradient Boosting (XGBoost) and Balanced Random Forest models, we demonstrate SMOTE's superior balance between accuracy (0.956) and fairness compliance (Statistical Parity Difference, ΔSPD=0.045; Equalized Odds Difference, ΔEOD=0.047). Key findings include: 1) SMOTE with XGBoost achieves the optimal balance between performance (0.956 accuracy) and fairness compliance (ΔSPD=0.045, ΔEOD=0.047), outperforming ADASYN, which suffers 45% lower F1-score despite marginally better fairness, and ROS in both accuracy and fairness; and 2) real-time monitoring reduces bias drift by 32% while maintaining sub-200ms latency.
Recommended Citation
Huang, Jenny and Turetken, Ozgur, "Mitigating Algorithmic Bias in Fraud Detection" (2025). Proceedings of the 2025 Pre-ICIS SIGDSA Symposium. 29.
https://aisel.aisnet.org/sigdsa2025/29