Enhanced Predictive Modeling for Anomaly Detection in Financial Transactions Using Machine Learning

Credit card fraud detection presents a significant challenge due to the extreme class imbalance in transaction datasets. Traditional machine learning models struggle to achieve high recall while maintaining precision, limiting their effectiveness in real-world applications. To address this issue, a systematic approach integrating IQR-based outlier removal, PowerTransformer normalization, Principal Component Analysis (PCA)-based feature selection, and advanced resampling techniques is proposed. Performance comparisons across multiple resampling methods identify Adaptive Synthetic Sampling (ADASYN) as the optimal choice, balancing precision and recall effectively. Ensemble models, including LightGBM, XGBoost, and CatBoost, are trained and optimized, achieving substantial performance gains over baseline models (ROC-AUC: 0.9132 → 0.9916). Further improvements are observed with Voting and Stacking ensemble strategies, leading to ROC-AUC 0.9947. Additionally, Stochastic Weight Averaging (SWA) is applied, enhancing the final performance to ROC-AUC 0.9970. The proposed framework demonstrates superior fraud detection capabilities by effectively handling data imbalance, feature selection, and model generalization, ensuring high detection accuracy while maintaining computational efficiency. These results contribute to the advancement of robust fraud prevention mechanisms in financial transactions. Index Terms—Fraud detection, machine learning, ensemble learning, class imbalance, SWA, ADASYN, PCA, LightGBM, XGBoost, CatBoost.