Incomplete multi-view clustering via virtual-label guided matrix factorization

Abstract

Multi-view clustering methods have achieved appealing performance by exploring the data from multiple sources. It often assumes that the multiple-view data are complete. However, such rigorous assumption does not be satisfied in many real-world applications, in which only some views of data are available. The existence of the incomplete-view data would disable the current multi-view clustering algorithms. To address this challenge, in this article, we present a novel virtual-label guided matrix factorization (VLMF) to facilitate the incomplete multi-view clustering. Specifically, in VLMF, we first employ the graph regularization of each view to mine the geometrical structure of data. Then, we utilize the virtual label to guide matrix factorization. Furthermore, the clustering process and the consensus latent representation learning of data are integrated into a joint framework. Experimental results on several incomplete multi-view datasets illustrate the validity of the proposed method.