<p>The classification of large-scale imbalanced data is a challenging task in machine learning. Ensemble methods based on undersampling are considered one of the most effective approaches to address this task. However, most existing undersampling ensemble methods are susceptible to inappropriate undersampling strategies, which may lead to the loss of useful information in the majority class, thereby affecting the model’s generalization ability. To tackle this issue, this paper proposes a novel dynamic self-paced undersampling ensemble (DSUE) method. This method first implements adaptive self-paced undersampling, which selects the most informative majority class samples based on their difficulty distribution to create balanced data subsets for each base classifier. Secondly, a sample weighting strategy is employed to identify critical samples in class-overlapping regions, thereby improving the model’s predictive performance and reducing the impact of noisy data. Finally, a weighted ensemble strategy based on the Gmean scores of base classifiers on the original dataset is introduced, which allocates higher weights to base classifiers that perform better on the original dataset in the final decision-making process. We conducted comparative experiments on 30 small-scale datasets and five large-scale datasets, comparing our method with nine related approaches. The experimental results demonstrate that the proposed DSUE method significantly outperforms the compared methods in evaluation metrics such as F1, Gmean, and AUC, particularly in large-scale imbalanced datasets, showing its superior effectiveness and competitiveness.</p>