Optimized Financial Fraud Detection using SMOTE-Enhanced Ensemble Learning with CatBoost and LightGBM

Fraud detection in banking transactions is critical due to the surge in digital transactions and associated fraudulent activities. For imbalanced datasets, conventional machine learning models yield low recall and high false-negative rates for minority (fraudulent) classes, limiting real-time fraud detection. This study evaluates ensemble learning strategies with SMOTE for balancing datasets and proposes a framework examining machine learning algorithms, including Logistic Regression, KNN, Decision Tree, Random Forest, XGBoost, CatBoost, and LightGBM. The SMOTE-enhanced ensemble method, using majority voting, achieves superior precision and recall, with ROC-AUC scores exceeding 0.99 for top models, offering insights for developing resilient, high-accuracy fraud detection systems.