Tensorized Anchor Graph Learning for Large-scale Multi-view Clustering

With the continuous development of information acquisition technologies, large-scale multi-view data increases rapidly. The enormous computational and storage complexity makes it very challenging to process these data in real-world applications. Most existing multi-view subspace clustering (MVSC) always suffers from quadratic space complexity and quadratic or even cubic time complexity, resulting in extreme limitations for large-scale tasks. Meanwhile, the original data usually contain lots of noise or redundant features, which further enhances the difficulty of the large-scale clustering tasks. This paper proposes a novel MVSC method for efficiently and effectively dealing with large-scale multi-view data, termed as tensorized anchor graph learning (TAGL) for large-scale multi-view clustering. Concretely, TAGL first projects the original multi-view data from the original space into the latent embedding space, where the view-consistent anchor matrix. Meanwhile, we establish the connection between the anchor matrix and the original data to construct multiple view-specific anchor graphs. Furthermore, these anchor graphs are stacked into a graph tensor to capture the high-order correlation. Finally, by developing an effective optimization algorithm, the high-quality anchors, anchor graph, and anchor graph tensor can be jointly learned in a mutually reinforcing way. Experimental results on several big sizes of datasets verify the superiority and validity of TAGL. Therefore, the proposed TAGL can efficiently and effectively handle large-scale data tasks for real-world applications.