Multi-view clustering on data with partial instances and clusters.

Most multi-view clustering algorithms apply to data with complete instances and clusters in the views. Recently, multi-view clustering on data with partial instances has been studied. In this paper, we study the more general version of the problem, i.e., multi-view clustering on data with partial instances and clusters in the views. We propose a non-negative matrix factorization (NMF) based algorithm. For the special case with partial instances, it introduces an instance-view-indicator matrix to indicate whether an instance exists in a view. Then, it maps the instances representing the same object to the same vector, and maps the instances representing different objects to different vectors. For the general case with partial instances and clusters, it further introduces a cluster-view-indicator matrix to indicate whether a cluster exists in a view. In each view, it also maps the instances representing the same object to the same vector, but it further makes the elements of the vector 0 if the elements correspond to missing clusters. Then it minimizes the disagreements between the approximated indicator vectors of instances representing the same object. Experimental results show that the proposed algorithm performs well on data with partial instances and clusters, and outperforms existing algorithms on data with partial instances.