ABM: Adaptive bias mitigation for class-imbalanced semi-supervised learning

Class-imbalanced semi-supervised learning (CISSL) poses significant challenges in real-world scenarios, where limited labeled data and skewed class distributions jointly hinder model generalization, especially for minority classes. Existing CISSL methods predominantly focus on adjusting model outputs, often overlooking the crucial role of representation learning in addressing classifier bias. In this paper, we propose a feature-aware adaptive bias mitigation framework that effectively alleviates class imbalance while enhancing representation learning. Specifically, our approach leverages the batch-wise feature mean as an additional input to guide the learning process, enabling the model to calibrate its representations against an unbiased reference. We refine classifier predictions by adjusting logits based on batch-wise feature mean, and further apply post-hoc logit adjustment to correct residual response bias. This combination not only improves pseudo-label quality but also fosters balanced and robust feature learning. Extensive experiments on four benchmark datasets under various class distribution settings demonstrate that our method consistently outperforms state-of-the-art competitors, achieving superior balanced accuracy and robustness against distribution mismatches. • We propose ABM, a novel adaptive bias mitigation framework tailored for class-imbalanced semi-supervised learning (CISSL). • ABM introduces a batch-wise feature mean mechanism that leverages context from labeled and unlabeled data to improve pseudo-label quality. • Extensive experiments demonstrate that ABM achieves superior generalization and bias reduction on CISSL benchmarks.