Abstract Class imbalance is a challenge that significantly hinders the performance of classification models in real-world applications. Existing techniques, including data-level resampling, cost-sensitive learning, and ensemble methods, often struggle to effectively address the simultaneous presence of these issues. The reasons include high computational costs and limited capacity to capture the structural characteristics of instance distributions. In this paper, we propose the Uncertainty-Aware Balanced Deep Forest (UABDF), a novel ensemble framework designed to enhance classification robustness and structural distribution of instances. UABDF integrates uncertainty estimation into a dynamic undersampling strategy guided by evidence-based learning, aiming to mitigate the influence of noise and class overlapping. To further improve generalization, UABDF introduces a dual-purpose Wasserstein-based alignment mechanism: Target Class Distribution Alignment ensures that the consensus prediction aligns closely with the true label distribution, and Cross-Classifier Distribution Alignment enforces consistency across individual classifier predictions within the ensemble. Extensive experiments on multiple benchmark datasets demonstrate that UABDF significantly improves learning performance in terms of F1-score and AUC, outperforming existing state-of-the-art methods in real-world imbalanced classification scenarios.
Gong et al. (Fri,) studied this question.