Enhancing Multiview Subspace Clustering with L_(1,2) Regularization and Self-Labeling Supervision

Abstract

In recent years, multiview subspace clustering has gained widespread attention due to its ability to effectively integrate complementary information from multiple views, revealing the underlying structure in high-dimensional data. However, existing methods still face challenges in handling complex data scenarios due to their limited representation power. Among these methods, the Multiview Deep Subspace Clustering Network (MvDSCN) has improved clustering performance to some extent by embedding multiview relationships into the feature learning and self-representation stages through the design of a diversity network (Dnet) and a universality network (Unet). However, we observe that the shared representation learned by MvDSCN lacks sufficient discriminative power, which negatively impacts the quality of the self-representation matrix. Furthermore, due to the limitations of its unsupervised learning strategy, the model struggles to effectively leverage latent label information to guide feature learning, thus constraining the improvement in clustering performance. To address these issues, we propose a novel multiview subspace clustering method, L12SL-MvSC, based on L_(1,2)regularization and self-labeling supervision. First, we apply regularization to the self-representation coefficient matrix to select discriminative sample relationships. Then, we introduce a self-labeling supervision strategy, which generates pseudo-labels to assist network training, further enhancing the quality of self-representation learning and clustering performance. Experimental results on benchmark datasets demonstrate the effectiveness of the proposed method.