Multi-granularity ensemble sample selection and label correction for classification with noisy labels

Abstract

Sample selection is crucial in classification tasks with noisy labels, yet most existing sample selection methods rely on a single criterion. These approaches often face challenges, including low purity of selected clean samples, and underfitting due to an insufficient number of selected clean training samples. To address these challenges, this paper proposes GNet-SSLC, a novel multi-granularity network framework that integrates multiple criteria ensemble sample selection (SS) and multiple views label correction (LC). In the SS phase, this paper proposes a metric learning-based dual k-Nearest Neighbor (k-NN) sample selection method. This method first uses corrected soft labels from the initial k-NN round to guide the selection of clean samples in the subsequent k-NN round. To further enhance selection accuracy, we combine this dual k-NN approach with a small loss sample selection technique through a voting mechanism. This multiple criteria ensemble method addresses the issues of low purity and instability inherent in single criterion approaches. In the LC phase, this paper designs a multiple views label correction framework that generates high-quality pseudo-labels for selected noisy samples. A key innovation of the framework is the design of a regularized contrastive learning loss, which optimizes the semi-supervised learning process by leveraging multiple views of training samples. The additional inclusion of training samples with high-quality pseudo-labels can effectively mitigate underfitting caused by a limited number of clean training samples. Experimental results on both synthetic and real-world noisy datasets indicate that GNet-SSLC enhances the purity and stability of the selected clean samples, and significantly improves classification performance. The enhancement is particularly notable with high noise rate dataset, such as CIFAR-100 dataset with 80% noise rate, achieving a 19.3% increase in classification accuracy compared to the baseline method.