Evaluating Machine Learning Algorithms for Effective Credit Card Fraud Detection

This study studies Machine Learning (ML) algorithms for tracing credit card fraud, including random forest, logistic regression, and Support Vector Machines (SVMs), using a dataset of 287,807 transactions from European cardholders. To resolve data imbalance, resampling was done after applying data preparation techniques. Ten balanced subsets of data were produced by undersampling the majority (non-fraud) class, each of which concentrated on a different aspect of the majority. These subsets were used to train individual models, and predictions were produced by majority voting. For comparison, models were trained using the original unbalanced data in addition to resampling. Accuracy, recall, precision, F1 scores, ROC-AVC scores, and a practical business assessment to determine their practical relevance were among the evaluation criteria. Logistic regression, trained on unbalanced data, achieved a ROC-AUC of 99.06% with transparent decision-making, outperforming models trained on balanced subsets, which, despite high AUC-ROC scores, suffered low precision and reduced practical reliability in fraud detection.