Incomplete multi-view spectral clustering

Abstract

Multi-view clustering algorithms mostly apply to data without incomplete instances. However, in real-world applications, representations for the same instance are probably absent from several but not all views. This incompleteness disables traditional multi-view clustering methods from grouping incomplete multi-view data. Recently, multi-view clustering methods on incomplete data have been proposed, and the existing methods have two limitations. One is that most methods were developed for incomplete datasets only with two views. The other is that most methods were incapable of grouping data with complex distributions. In this paper, we propose a novel incomplete multi-view clustering algorithm named IMSVC, in which we adopt spectral analysis to supervise the common representation extracted from all the views. Firstly, IMVSC constructs a bipartite graph for each view. By introducing an instance-view indicator matrix to indicate whether a representation exists in a view or not, we calculate the edge weights of bipartite graph based on the point-to-point similarity. Secondly, IMVSC constructs the multi-view relationship by guiding the multiple views to share the same instance partitioning. Finally, we create a novel iterative method to optimize IMVSC. Experimental results show sound performance of the proposed algorithm on several incomplete datasets.