Multi-view Intact Space Clustering

Abstract

Multi-view clustering is a hot research topic due to the urgent need for analyzing a vast amount of heterogeneous data. Although many multi-view clustering methods have been developed, they have not addressed the view-insufficiency issue. That is, most of the existing multi-view clustering methods assume that each individual view is sufficient for constructing the cluster structure, which is however not guaranteed in real applications. In this paper, we propose a novel multi-view clustering method termed multi-view intact space clustering (MVIC), which is able to simultaneously recover the latent intact space from multiple insufficient views and construct the cluster structure from the resulting intact space. For each view, a view generation function is designed to map the latent intact space representation into the view representation. Since we are given the view representation, by mapping back from each individual view representation, the latent intact space can be restored, based on which the matrix factorization based clustering can be applied. Therefore, the proposed model is composed of two components, namely the reconstruction error of the latent intact space and the distortion error of data clustering in intact space. An alternating iterative method is designed to solve the optimization of the model. Experimental results conducted on a wide-range of multi-view datasets have confirmed the superiority of our method over state-of-the-art approaches.