Dual-space learning with low-rank and sparse constraints for multi-view subspace clustering

Multi-view subspace clustering (MSC) has attracted considerable attention because it integrates consistent and complementary information from multiple views. However, many existing methods learn different types of structural information within a single representation space, which may not clearly distinguish their respective roles. First, representing heterogeneous information from different views within a single space may make it difficult to preserve both global consistency across views and local structures within individual views. Second, although low-rank and sparse regularizations have been widely used in subspace clustering, their roles in different latent representation spaces have received less attention. To address these issues, we propose Dual-Space Learning with Low-Rank and Sparse Constraints for Multi-View Subspace Clustering (DSLSMSC). Specifically, the proposed method simultaneously learns latent representations from the joint-view space and the original-view spaces. Low-rank regularization is imposed on the joint-view representation to capture the shared structure across views, while sparse regularization is imposed on the original-view representations to preserve view-dependent local relationships. The learned representations are further integrated via tensor nuclear norm (TNN) regularization to exploit high-order correlations across spaces. Experimental results on multiple benchmark datasets demonstrate the effectiveness of DSLSMSC.