Hybrid Reinforcement Learning-Active Learning Framework for Real-Data Augmentation in Imbalanced Credit Scoring

Class imbalance is a critical challenge in credit scoring, where the dominance of majority class samples reduces predictive performance for minority instances. Traditional methods, such as SMOTE and random undersampling, attempt to rebalance datasets but often introduce synthetic noise or discard valuable data. This paper introduces Augmentation Based on Uncertainty and Difficulty (UDDA), a novel real-data augmentation framework that avoids both synthetic data generation and majority class removal. UDDA leverages a hybrid of Reinforcement Learning (RL) and Active Learning (AL) to identify and prioritize real, informative, and difficult-to-classify samples. This approach enhances model robustness while preserving dataset integrity. Experimental evaluation on twenty benchmark imbalanced tabular datasets demonstrates that UDDA outperforms established methods, including SMOTE and undersampling, across key metrics such as precision, recall, F1-score, and accuracy. UDDA sets a new direction in imbalanced learning by offering a practical, interpretable, and effective solution for improving classification performance in credit scoring applications.