Imbalanced Multi-View Semi-Supervised Classification via Label-Level Contrastive Learning and Pseudo-Label Refinement
Xinchao Lu, Lihua Zhou · 2025
Due to the expensive cost of labeling, semi-supervised classification on multi-view data with limited labels is inevitable. However, existing multi-view semi-supervised classification algorithms often overlook the fact that training class distributions in practical applications are imbalanced, which results in poor predictions for minority classes. To address this challenge, we propose a multi-view semi-supervised classification algorithm based on label-level contrastive learning and pseudo-label refinement to mitigate prediction bias. Our algorithm first leverages deep autoencoders to extract the underlying features of samples and introduces label-level contrastive learning within each view to further enhance the discriminative ability of sample features. Then, a pseudo-label refinement component is designed to adaptively adjust the predicted class distribution, improving the model’s perception of minority classes. Experimental comparisons with state-of-the-art algorithms validate the effectiveness of the proposed method.