Beyond High-Confidence Sampling: Boosting Data Utilization and Model Calibration in Semi-Supervised Learning
IEEE Access · 2026 · 14 · 46429-46441
DOI: 10.1109/access.2026.3677254 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shilei Tang und 1 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Semi-supervised learning (SSL) aims to improve model performance by leveraging abundant unlabeled data in conjunction with limited labeled data. Most existing SSL methods rely on pseudo-labeling approaches, which focus almost exclusively on high-confidence samples (i.e., pseudo-labeled samples) while neglecting low-confidence ones. This strategy suffers from two major limitations. First, prevailing SSL models tend to generate overconfident predictions, allowing noisy samples to pass through the pseudo-labeling process and contaminate the training procedure, thereby reducing model performance. Second, to reduce noise, these methods typically apply high-confidence thresholds to filter out low-confidence unlabeled samples, leading to substantial underutilization of available data. To address these issues, we propose a class-aware margin (CAM) constraint, which regulates the logit distances between model predictions across different classes within reasonable bounds, thus improving model calibration. To increase the utilization of unlabeled data, we further introduce a negative class penalty (NCP), which applies negative learning to samples whose confidence falls below the threshold. This design enables the model to effectively exploit all unlabeled data. We evaluate the proposed methods on several image classification benchmarks under both balanced and imbalanced settings. Experimental results demonstrate that our approach significantly enhances both the discriminative power and calibration capability of the trained models. Code is available at: https://github.com/wpsapp123/CAM-NCP
Zitieren
APA
Tang, S., & Qi, Y. (2026). Beyond High-Confidence Sampling: Boosting Data Utilization and Model Calibration in Semi-Supervised Learning. IEEE Access, 14, 46429-46441. https://doi.org/10.1109/access.2026.3677254
Vancouver
Tang S, Qi Y. Beyond High-Confidence Sampling: Boosting Data Utilization and Model Calibration in Semi-Supervised Learning. IEEE Access. 2026;14:46429-46441. doi:10.1109/access.2026.3677254
BibTeX
@article{Tang2026,
author = {Tang, Shilei and Qi, Yunsong},
title = {Beyond High-Confidence Sampling: Boosting Data Utilization and Model Calibration in Semi-Supervised Learning},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {46429--46441},
doi = {10.1109/access.2026.3677254}
}
EUVIMED