Learning Imbalanced Data with Beneficial Label Noise

Guangzheng Hu, Feng Liu, Mingming Gong, Guanghui Wang, Liuhua Peng

Keywords: Imbalanced learning, beneficial label noise, classificaition accuracy

Abstract: Data imbalance and label noise are common factors hindering the classifier's performance. Data-level approaches to addressing imbalanced learning usuallyinvolve resampling by adding or removing samples, which often results in information loss or generative errors. Building upon theoretical studies of the impact of imbalance ratio on decision boundaries across various evaluation metrics in binary classification, it is uncovered that introducing appropriate label noise can alter the biased decision boundaries and thus enhance the performance of classifiers in imbalanced learning. In this paper, we introduce the Label-Noise-based Re-balancing (LNR) approach to solve both binary and multi-class imbalanced classifications by employing a novel design of asymmetric label noise model. In contrast to other data-level methods, our approach is easy to implement and alleviates the issues of informative loss and generative errors. We validated the superiority of this method on synthetic and real-world datasets. More importantly, our LNR approach can integrate seamlessly with any classifiers and other algorithm-level methods for imbalanced learning. Overall, our work opens up a new avenue for addressing imbalanced learning, highlighting the potential advantages of balancing data through beneficial label noise.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9243