A Dual Correction Guarantee Mechanism for Numerical Label Noise
IEEE Transactions on Knowledge and Data Engineering · 2026-09 · 38 (9) · 5567-5580
DOI: 10.1109/tkde.2026.3708662 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Knowledge and Data Engineering, 2026-09. Verfasst von Yaqing Guo und 6 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Fundamental Research Program of Shanxi Province.
Abstract (Originalsprache)
Label noise usually leads to the degradation of the model generalization in supervised learning, especially for deep neural networks. Although substantial studies have explored learning from noisy labels, most of them focus on classification, with limited effectiveness on noisy regression tasks. Moreover, advanced techniques for addressing numerical label noise typically discard almost all noisy examples directly or downweight them, thus resulting in a partial exploration of training data. In this work, we propose a numerical label noise correction method under a dual correction guarantee mechanism (named DCGM), aiming to address the problem of sample information loss and effectively reduce the noise level. First, we theoretically prove that DCGM can reduce the generalization error bound for noisy regression. And then, we resist label noise influence under a dual correction guarantee mechanism. (1) We determine genuine neighborhoods through feature embedding representations, rather than directly selecting nearest neighbors in the original feature space, to identify noisy labels more precisely and provide usable and accurate sample information for label correction. Then, we define the predictive uncertainty, and only those noisy labels with low uncertainty values are selectively corrected and retained. Other noisy samples are discarded. (2) We construct a robust heteroscedastic loss for performing post-correction. Samples with false corrections and unidentified noisy samples will be assigned smaller weights during the updating of network parameters, to implicitly conduct the second round of label correction. Experimental results demonstrate that DCGM can achieve state-of-the-art performance on noisy regression tasks.
Förderung
- National Natural Science Foundation of China
- Fundamental Research Program of Shanxi Province
Zitieren
APA
Guo, Y., Cui, L., Jiang, G., Lei, F., Hou, S., Xu, H., & Wang, W. (2026). A Dual Correction Guarantee Mechanism for Numerical Label Noise. IEEE Transactions on Knowledge and Data Engineering, 38(9), 5567-5580. https://doi.org/10.1109/tkde.2026.3708662
Vancouver
Guo Y, Cui L, Jiang G, Lei F, Hou S, Xu H, et al. A Dual Correction Guarantee Mechanism for Numerical Label Noise. IEEE Transactions on Knowledge and Data Engineering. 2026;38(9):5567-5580. doi:10.1109/tkde.2026.3708662
BibTeX
@article{Guo2026,
author = {Guo, Yaqing and Cui, Lingxuan and Jiang, Gaoxia and Lei, Fan and Hou, Senyu and Xu, Hang and Wang, Wenjian},
title = {A Dual Correction Guarantee Mechanism for Numerical Label Noise},
journal = {IEEE Transactions on Knowledge and Data Engineering},
year = {2026},
volume = {38},
number = {9},
pages = {5567--5580},
doi = {10.1109/tkde.2026.3708662}
}
EUVIMED