Comprehensive Multi-View Clustering via Low-Dimensional Subspace Representation

Although previous multi-view subspace clustering approaches have achieved tremendous progress, there still exist some weaknesses in most of them. Firstly, previous algorithms usually adopt self-representation to construct the loss function, which often leads to high computational complexity. Secondly, they typically learn similarity matrices without block structures, which may adversely affect the clustering partition. Thirdly, the view weights are often assigned equally, neglecting that each view feature descriptor exerts different degrees of influence on clustering results. In light of this, we design a new clustering model with dynamic subspace representation and propose a novel Multi-view Subspace Clustering method based on Anchor Graph and Consensus Representation (MSCAGCR). Specifically, representative samples are selected and iteratively updated to reduce the time cost. Moreover, the introduction of a rank constraint on the Laplacian matrix corresponding to the similarity graph ensures that the clustering results contain connected components whose number is identical to the number of classes. Furthermore, the exponential weighting strategy enhances the usability of high-quality view features while reducing the impact of inferior ones. Extensive experiments conducted on eight benchmark datasets demonstrate that MSCAGCR outperforms existing methods in terms of clustering effectiveness and efficiency.