Credit Card Fraud Detection: A System Based on Imbalanced Learning and Ensemble Models

Credit card fraud detection is a critical challenge in financial security, exacerbated by the highly imbalanced nature of transaction datasets where fraudulent cases are rare. This paper proposes a system leveraging SMOTE for oversampling and XGBoost as an ensemble model, with a minor modification incorporating scale pos weight to handle class imbalance effectively. Our approach optimizes feature weights while ensuring computational efficiency on standard CPUs. Experiments on the Kaggle credit card fraud dataset (approximately 150MB) demonstrate superior performance, achieving an AUC of 0.977 and F1 score of 0.892 for the proposed model, outperforming baselines like Logistic Regression and Random Forest. Key visualizations include ROC and Precision-Recall curves showing enhanced minority class detection, confusion matrices highlighting low false positives, feature importance rankings emphasizing variables like V14 and V17, and violin plots illustrating amount distributions by class. Ablation studies confirm the necessity of SMOTE and key features, with detailed metrics in tables. The method’s robustness is validated through cross-validation, with training times under 30 s per model. This work contributes a practical, high-performing solution for real-time fraud detection, suitable for EI conference scales with 6-7 pages. Future extensions could integrate real-time streaming data.