MetaMatch: Robust Self-Training Under Label Noise
Year: 2026
Type: article
Abstract: As an elegant design of semi-supervised learning (SSL) algorithms, the self-training method can achieve superior performance when dealing with unlabeled data. Under certain conditions, these methods generate pseudo-labels on unlabeled data and train the model with the newly labeled data. The pseudo-labeling process can assign model confidence to unlabeled data near the decision boundary and expand the boundary in regions of low density. Consequently, the method achieves more precise contour and demonstrate high efficiency in various real-world applications. In certain practical scenarios, however, labeled data is always biased or corrupted by label noise. Our ablation study confirms that label noise can detrimentally degrade the performance of self-training by destroying the consistency of the pseudo-labeling process. To address this issue, we propose a novel robust self-training framework through consistency regularization and sample re-weighting to mitigate the label noise impact on both supervised learning and unsupervised learning part. Based on our experiments on synthetic and real human label noise datasets, our proposed method performs robustly throughout self-training in the presence of label noise without requiring any clean data. The performance is also comparable with other state-of-the-art SSL methods that can handle label noise circumstances.
Authors: Xihe Qiu, Yuxuan Fu, Shaojie Shi, Chao Qu, Xiaoyu Tan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204148313