.png){kind=icon}
Return
MBP: Rethinking class-imbalanced semi-supervised learning from the multi-binary perspective
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.133273.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Class-Imbalanced Semi-Supervised Learning (CI-SSL) plays a critical role in real-world applications. However, conventional methods often generate majority-class-biased pseudo-labels due to skewed data distributions, which degrade performance on minority classes. We identify that the core bottleneck of CI-SSL resides in the failure of the global optimization paradigm in traditional multi-class frameworks to adapt to class-specific requirements: majority classes dominate the optimization process and neglect minority classes, while existing unified correction strategies fail to achieve precise class-level bias mitigation. To address this challenge, we propose MBP (Multi-Binary Perspective), a novel approach that reformulates CI-SSL from a multi-binary classification lens. MBP decomposes the multi-class imbalanced task into K one-vs-all (OVA) binary subtasks, enabling targeted optimization tailored to extreme class imbalance. Within this framework, we design an Adjusted Loss (ADL) integrated with a Dual Adjustment Strategy (DAS) to safeguard minority-class representations. Additionally, a Class-Wise Dynamic Threshold (CWDT) is introduced to adaptively set thresholds for generating high-quality pseudo-labels. Trained within a dual-branch architecture that combines linear and multi-binary classifiers, MBP achieves State-of-the-Art (SOTA) performance on the CIFAR-10-LT, CIFAR-100-LT, and STL-10-LT datasets. This work provides a lightweight yet robust solution for CI-SSL while offering a new paradigm for mitigating pseudo-label bias.
Keywords:
Class-Imbalanced Semi-Supervised Learning
Multi-Binary Perspective
Pseudo-Label Bias
Minority-Class Optimization
Dual Adjustment Strategy
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
Cited Papers
No cited papers available