Learning with Imbalance Noisy Labels via Confidence-guided Sample Mixing and Negative Learning

Deep learning methods have achieved significant performance improvements by relying on large-scale annotated data. However, real-world datasets are often affected by two major issues. The first is label noise caused by annotation errors, and the second is class imbalance resulting from significant disparities in the number of samples across different classes. The combined effect of the two issues can severely undermine model robustness. In real-world settings where label noise coexists with class imbalance, conventional denoising methods often prove inadequate, as they are predicated on a balanced data distribution. For instance, a strategy that select samples based on low loss struggle to distinguish whether a high-loss sample belongs to a hard-to-learn minority class or is indeed a mislabeled instance. To address the above challenge, we propose a sample selection method with sample mixing and negative learning. The core of this method first utilizes the confidence-guided sample mixing, which dynamically regulates mixing ratios based on the confidence of the model in the samples, thereby enhancing the reliability of clean samples during training. Second, we design a strategy of the negative learning based on the average confidence margin to re-utilize low-quality noisy samples. In addition, we also introduce the L1 \& L2 regularization to mitigate model overfitting. Experimental results on the CIFAR-10 and CIFAR-100 datasets demonstrate that our method outperforms state-of-the-art competitors under various scenarios where class imbalance and label noise coexist.