Kernel-space oversampling and adaptive ensemble learning for robust imbalanced data classification

Abstract This paper proposes a robust fuzzy kernel-based possibilistic clustering-means adaptive weighted random forest (FKAW-RF) framework for imbalanced data classification that integrates kernel-space oversampling with adaptive ensemble learning. The method first maps the original dataset into a high-dimensional kernel space, where a fuzzy kernel-based possibilistic clustering-means (FKPCM) algorithm is employed to identify and remove boundary samples, thereby reducing the influence of noise and class overlap. Within this kernel space, a kernel-conventional synthetic minority oversampling technique (KSMOTE) strategy is applied to perform high-quality oversampling of minority class instances, and the generated samples are projected back into the input space through preimage estimation. To further enhance classification robustness, an adaptive weighted random forest (AW-RF) is constructed as the ensemble classifier, in which underperforming or redundant decision trees are pruned based on feature correlation analysis, and the remaining trees are adaptively weighted using dual penalty factors reflecting both local and global error rates. Extensive experiments on UCI, KEEL, and Tennessee-eastman (TE) benchmark datasets demonstrate that the proposed kernel-space oversampling and adaptive ensemble learning approach consistently outperforms conventional synthetic minority oversampling technique (SMOTE) variants and state-of-the-art classifiers across multiple evaluation metrics, including accuracy, precision, recall, and G-mean, confirming its effectiveness and robustness for challenging imbalanced classification tasks.