Comparing Machine Learning Models and Voting Ensembles for Credit Card Fraud Detection

Credit card fraud poses a growing threat to global financial systems. The rarity of fraudulent transactions makes this a highly imbalanced classification problem, requiring models that can maintain high recall without sacrificing precision. The effectiveness of four supervised learning models is assessed in this study—Logistic Regression, Random Forest, eXtreme Gradient Boosting (XGBoost), and a Soft Voting Ensemble—on a PCA-transformed credit card dataset with a 5% fraud ratio. Area Under the Curve (AUC), F1-score, precision, and recall are evaluation measures. Logistic Regression demonstrated high recall but poor overall balance, leading to its exclusion from test set evaluation. Random Forest achieved perfect precision but lower recall, while XGBoost and the Soft Voting Ensemble showed stronger balance across metrics. Soft Voting produced the best F1-score and most stable performance across both validation and test sets. These results indicate that ensemble methods, particularly soft voting, can effectively address imbalanced classification in fraud detection. Future work may explore alternative sampling strategies, larger datasets, and model tuning frameworks to further improve detection performance and adaptability to real-world scenarios.