Comparing Ensemble Methods for Credit Card Fraud Detection: A Performance Analysis on Multiple Datasets

Credit card fraud detection has become increasingly crucial as digital payment systems expand globally. This paper presents a comprehensive comparative analysis of ensemble learning methods for credit card fraud detection across multiple datasets. We systematically evaluate various ensemble approaches including Random Forest, XGBoost, LightGBM, stacking methods, and hybrid approaches, analyzing their performance on the widely-used IEEE-CIS dataset and other benchmark datasets. Our experimental evaluation demonstrates that stacking ensemble methods achieve superior performance with AUC-ROC scores up to 0.943, while maintaining computational efficiency suitable for real-time deployment. We analyze the impact of data balancing techniques, feature engineering strategies, and explainable AI integration on ensemble performance. The results show that ensemble methods consistently outperform individual classifiers, with stacking approaches providing the best balance between accuracy and interpretability for practical fraud detection systems.