Self-Paced Pseudo Label Refinement for Multi-view Subspace Clustering

Abstract

Multi-view clustering can capture common representations from multi-view data that contain complementary information of different views, which has been applied in many fields, such as computer vision, natural language processing, medicine. As one of the most popular attractive directions, multi-view subspace clustering focuses on learning common subspace representations of different views that can reveal the underlying clustering structure of multi-view data. However, for learning more discriminative common subspace representations, current multi-view subspace clustering methods utilize pseudo labels of all samples to train the self-supervised model, which neglects the pseudo label noise problem in the self-supervised multi-view subspace clustering. To address the challenge, we propose a novel paradigm termed self-paced pseudo label refinement paradigm, which aims at dynamically refining pseudo labels in the self-supervised multi-view subspace clustering. First, we design the self-paced learning strategy to prevent noisy pseudo labels from interfering with the self-supervised model training and improve the discrimination of view-specific semantic features. Then, we develop the adaptive update mechanism to learn weights of different views in an adaptive way and combine view-specific semantic features with obtained weights to refine pseudo labels. Finally, to implement multi-view subspace clustering based on the proposed paradigm, the deep self-supervised multi-view subspace clustering network is developed. Experiments on six clustering benchmarks verify the proposed approach outper-forms state-of-the-art multi-view subspace clustering methods.