Multi-view Spectral Clustering via Multi-view Weighted Consensus and Matrix-Decomposition Based Discretization

Abstract

In recent years, multi-view clustering has been widely used in many areas. As an important category of multi-view clustering, multi-view spectral clustering has recently shown promising advantages in partitioning clusters of arbitrary shapes. Despite significant success, there are still two challenging issues in multi-view spectral clustering, i.e., (i) how to learn a similarity matrix for multiple weighted views and (ii) how to learn a robust discrete clustering result from the (continuous) eigenvector domain. To simultaneously tackle these two issues, this paper proposes a unified spectral clustering approach based on multi-view weighted consensus and matrix-decomposition based discretization. In particular, a multi-view consensus similarity matrix is first learned with the different views weighted w.r.t. their confidence. Then the eigen-decomposition is performed on the similarity matrix and a set of c eigenvectors are obtained. From the eigenvectors, we first learn a continuous cluster label and then discretize it to build the final clustering label, which avoids the potential instability of the conventional k-means discretization. Extensive experiments have been conducted on multiple multi-view datasets to validate the superiority of our proposed approach.