A Comparative Study of Machine Learning Methods for Credit Card Fraud Detection

Credit card fraud detection represents a critical challenge in financial security, characterized by extreme class imbalance and evolving fraudulent patterns. This paper presents a comprehensive comparative analysis of eight machine learning algorithms for credit card fraud detection using a real-world anonymized transaction dataset. We implement a complete pipeline featuring RobustScaler preprocessing for outlier resistance and strategic class balancing. Our evaluation encompasses Logistic Regression (with L1 and L2 regularization), k-Nearest Neighbors, Decision Trees (employing both Gini impurity and entropy criteria), Random Forest, XGBoost, and Support Vector Machines. Experimental results demonstrate that Support Vector Machines achieve the most consistent performance with the highest cross-validation ROC-AUC (0.9841 ± 0.0032), while XGBoost attains the best test set performance (ROC-AUC: 0.9877). We observe distinct precision-recall trade-offs: SVM maximizes precision (0.99), Logistic Regression maximizes recall (0.96), while ensemble methods provide balanced performance. The complete implementation is publicly available to ensure reproducibility and serve as a benchmark for future research.