Machine Learning Approaches for Credit Card Fraud Detection‎in Severely Imbalanced Datasets: A Comparative Analysisof ‎Classification and Anomaly Detection Methods

Abstract

Credit card fraud presents a persistent threat to financial institutions, exacerbated by the rise of digital payments and the complexity of ‎fraudulent schemes. This study investigates machine learning (ML) approaches for fraud detection in severely imbalanced datasets, focusing ‎on three key objectives: comparing classification and anomaly detection models under extreme class imbalance, identifying transaction ‎features with the highest discriminative power, and optimizing decision thresholds using cost-sensitive evaluation to minimize business ‎impact. Utilizing a dataset of 999 transactions with a fraud rate of 0.2% (498.5:1 imbalance), we implemented supervised methods (logistic ‎regression, random forest, gradient boosting) and unsupervised anomaly detection (Isolation Forest, One-Class SVM, Local Outlier ‎Factor). Results show that ensemble-based models, particularly Gradient Boosting, achieved superior performance (AUC-ROC = 0.956; ‎AUC-PR = 0.378) with perfect recall and improved precision relative to other methods. Feature analysis identified anonymized PCA-‎derived variables (V14, V10, V12) as the most discriminative indicators of fraudulent activity. Threshold optimization at 0.9 minimized ‎operational costs ($2,985) while maintaining full recall, yielding an estimated annual net benefit of $68,985 and a return on investment of ‎‎186.7%. This study contributes to the literature by integrating algorithm benchmarking, feature importance evaluation, and cost-sensitive ‎threshold optimization in an end-to-end fraud detection framework. The findings underscore the importance of ensemble learning, ‎imbalanced evaluation metrics (AUC-PR, precision, recall), and business-driven threshold calibration for developing effective and ‎economically viable fraud prevention systems. Future research should explore larger datasets, adaptive learning to address concept drift, and ‎explainable AI techniques to enhance interpretability and regulatory compliance‎.