Improvement of Bank Fraud Detection Through Synthetic Data Generation with Gaussian Noise
Fray L. Becerra-Suarez, Halyn Alvarez-Vasquez, Manuel G. Forero
Abstract
Bank fraud detection faces critical challenges in imbalanced datasets, where fraudulent transactions are rare, severely impairing model generalization. This study proposes a Gaussian noise-based augmentation method to address class imbalance, contrasting it with SMOTE and ADASYN. By injecting controlled perturbations into the minority class, our approach mitigates overfitting risks inherent in interpolation-based techniques. Five classifiers, including XGBoost and a convolutional neural network (CNN), were evaluated on augmented datasets. XGBoost achieved superior performance with Gaussian noise-augmented data (accuracy: 0.999507, AUC: 0.999506), outperforming SMOTE and ADASYN. These results underscore Gaussian noise’s efficacy in enhancing fraud detection accuracy, offering a robust alternative to conventional oversampling methods. Our findings emphasize the pivotal role of augmentation strategies in optimizing classifier performance for imbalanced financial data.