A Comparative Evaluation of Class Imbalance Handling Techniques for Credit Card Fraud Detection

Credit card fraud detection presents a challenging classification task due to extreme class imbalance, where fraudulent transactions constitute less than 1% of all observations. Selecting an appropriate imbalance handling technique is critical, yet the comparative performance of these techniques under varying imbalance severities remains insufficiently understood. This study conducts a systematic empirical evaluation of nine class imbalance handling techniques across two publicly available fraud detection datasets exhibiting different imbalance ratios (578:1 and 90:1). The techniques evaluated span data-level resampling (SMOTE, ADASYN, Borderline-SMOTE, SMOTE combined with Edited Nearest Neighbors, and random undersampling), algorithm-level cost-sensitive learning (class weighting), and ensemble-based approaches (EasyEnsemble, RUSBoost, and Balanced Random Forest). Each technique is paired with four base classifiers---Logistic Regression, Random Forest, XGBoost, and LightGBM---and assessed using five evaluation metrics: AUC-ROC, PR-AUC, F1-score, Matthews Correlation Coefficient, and recall. Results indicate that ensemble-based methods, particularly EasyEnsemble, achieve the most consistent improvements across both datasets. Hybrid resampling via SMOTE with Edited Nearest Neighbors produces comparable gains among data-level methods. A notable finding is that standard SMOTE, while improving AUC-ROC and F1-score, can reduce PR-AUC relative to the untreated baseline under severe imbalance. Cost-sensitive class weighting emerges as a computationally efficient alternative that preserves strong PR-AUC performance. These findings provide practical guidance for practitioners selecting imbalance handling strategies in fraud detection applications.