Incomplete multi-view clustering with partially mapped instances and clusters

Abstract Most multi-view clustering methods assume that each view has complete instances and clusters. However, in real world applications, the instances or clusters may be missed in some views. Recently, multi-view clustering on data with partially mapped instances has been studied. In this paper, we study the multi-view clustering on data with partially mapped instances and clusters to extend the application of multi-view clustering. We propose a NMF (Non-negative Matrix Factorization) based algorithm which separately deals with the mapped clusters/instances and the individual clusters/instances, i.e., both the basis matrix and the indicator matrix consist of a mapped part and an individual part. By bounding the mapped instances to reduce to the same indicator vectors, the mapped instances and clusters connect multiple views and guide to find the indicator vectors of all the instances. Furthermore, we improve the algorithm by using locally geometrical information to reduce the negative impact caused by multi-view interaction. Experiments show that the proposed algorithms perform well on data with partially mapped instances and clusters.