Learning the consensus and complementary information for large-scale multi-view clustering

Abstract

The multi-view data clustering has attracted much interest from researchers, and the large-scale multi-view clustering has many important applications and significant research value. In this article, we fully make use of the consensus and complementary information, and exploit a bipartite graph to depict the duality relationship between original points and anchor points. To be specific, representative anchor points are selected for each view to construct corresponding anchor representation matrices, and all views' anchor points are utilized to construct a common representation matrix. Using anchor points also reduces the computation complexity. Next, the bipartite graph is built by fusing these representation matrices, and a Laplacian rank constraint is enforced on the bipartite graph. This will make the bipartite graph have k connected components to obtain accurate clustering labels, where the bipartite graph is specifically designed for a large-scale dataset problem. In addition, the anchor points are also updated by dictionary learning. The experimental results on the four benchmark image processing datasets have demonstrated superior performance of the proposed large-scale multi-view clustering algorithm over other state-of-the-art multi-view clustering algorithms.

Keywords: Bipartite graph; Complementarity; Consensus; Multi-view clustering.