Conservative plausibility-filtered SMOTE for credit card fraud detection under extreme class imbalance

Detecting credit card fraud is a complex tabular learning problem due to the underrepresentation of fraudulent events, extreme class imbalance, and scattered minority areas. To address this limitation, a conservative minority synthesis strategy is proposed, combining the removal of minority outliers, synthetic data generation using SMOTE, geometric filtering, and supervised plausibility filtering. Unlike conventional oversampling, the method does not seek to artificially balance classes but rather to retain only locally consistent and discriminatively useful synthetic samples. The approach was evaluated on the ULB Credit Card Fraud Detection and IEEE-CIS Fraud Detection datasets using a 60/20/20 stratified partitioning, preventing information leakage between training, calibration, and testing. On ULB, the proposal retained 638 final synthetic samples, raising the fraud rate of the augmented dataset to 0.005396. In IEEE-CIS, 947 final synthetic samples were retained, with a final fraud rate of 0.037563. Compared to techniques such as SMOTE, Borderline-SMOTE, ADASYN, and SimpleVAE, the proposed method showed competitive performance, achieving the highest AUPRC in four of six classifiers in ULB and in three of six classifiers in IEEE-CIS. Ablation analyses and statistical validation show that supervised plausibility filtering helped control the quality of the generated samples, although its benefit is not uniform and depends on the classifier and dataset used.