.png){kind=icon}
.png){kind=icon}
Return
Semi-supervised local entropy-decayed oversampling for imbalanced data
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.115009.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Real-world classification tasks often suffer from class imbalance, which poses significant challenges to model performance. Recent studies suggest that the degradation in classification accuracy stems not merely from the imbalance itself, but from underlying complex data characteristics such as class overlap, small disjuncts, sparse distributions, and outliers. However, most existing oversampling methods address only a subset of these challenges and lack the ability to jointly handle multiple structural complexities. To this end, we propose Semi-Supervised Local Entropy-Decayed Oversampling (SLEO), a novel method that integrates semi-supervised clustering with entropy-guided mechanisms to jointly address class overlap, small disjuncts, and outliers, while also alleviating the adverse effects of sparse distributions. Specifically, SLEO employs label-guided semi-supervised hierarchical clustering to model the minority class structure and to identify complex sub-concepts such as boundaries, small disjuncts, sparse regions and outliers. A local entropy-decay mechanism is then used to detect and clean high-uncertainty observations in overlapping regions. Finally, an entropy-weighted hypersphere-based observation generation strategy is applied within each subcluster, ensuring that synthetic observations are distributed in more discriminative and structurally stable regions. Extensive experiments on 65 benchmark imbalanced datasets demonstrate that SLEO significantly outperforms 13 state-of-the-art sampling techniques across multiple classifiers and evaluation metrics, as evidenced by average rankings, performance distribution analyses, and Holm’s post-hoc tests. Furthermore, correlation analysis between various data complexity characteristics and classification performance highlights SLEO’s superior adaptability and robustness.
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W