Integrating Class Imbalance Solutions into Fraud Detection Systems: A Systematic Literature Review

This paper reviews the three primary methodologies to tackle class imbalance: data-level techniques, algorithmic approaches, and hybrid methods. To guarantee a more balanced dataset, data-level techniques like sampling and oversampling with SMOTE concentrate on modifying the data before entering the model. Weighting and cost-sensitive learning algorithms change the learning process to minimize bias toward the majority class. Hybrid approaches blend these strategies with advanced machine learning techniques like deep learning to create robust models that efficiently handle skewed data distributions. The study explores various innovative solutions like Dynamic Weighted Entropy and SMOTified-GAN. These solutions utilize deep learning frameworks and generative adversarial networks to generate realistic and balanced data distributions. This review also examines the evaluation methods used to determine the effectiveness of these techniques, emphasizing metrics like the F1 Score, AUC, and ROC Curve, which are essential for assessing models in imbalanced datasets. The findings from this review reveal that while numerous studies have explored different aspects of class imbalance, significant gaps still exist in effectively integrating these solutions into practical fraud detection systems. The study provides critical insights for enhancing the robustness and accuracy of these systems and identifies areas requiring further research and development to refine these integration techniques. This contribution is vital for developing improved policies and industry standards in digital fraud management, aiming to bolster the operational sustainability of digital business environments..