Advanced Detection of Credit Card Fraud Using Ensemble Methods on Skewed and Uniform Datasets

Credit card fraud has become a major challenge with the rapid growth of online financial transactions and digital payment systems. The increasing volume of transaction data and the highly imbalanced nature of fraud datasets make accurate detection difficult for traditional methods. This study evaluates the effectiveness of a number of modern classifiers on both balanced datasets and imbalanced datasets using oversampling techniques like SMOTE, RandomOverSampler, and ADASYN, including Random Forest (RF), Gradient Boosting (GB), XGBoost (XGB), LightGBM (LGBM), CatBoost (CatB), and a proposed ensemble voting framework. For imbalanced data, SMOTE provided better F1-score and Kappa scores by generating synthetic minority samples, while ADASYN improved recall but occasionally introduced noisy points, which reduced precision. The proposed ensemble voting framework, which combined RF, CatB, and LGBM, provided superior stability, precision, and AUC scores, outperforming the remaining models. Strong individual performance was continuously generated by boosting models such as RF, LGBM, and CatB. All classifiers performed almost perfectly on balanced datasets with F1 scores exceeding 99.9\% and Kappa scores above 99.95\%. The ensemble approach consistently produced superior outcomes by using the complementary advantages of bagging and boosting methods to reduce variance, improve generalization, and maintain robustness across various training sizes. The findings indicate that a reliable and efficient framework for credit card fraud detection was established through the integration of advanced boosting and bagging techniques in both imbalanced and balanced datasets.