Semi-Paired Multiview Clustering Based on Nonnegative Matrix Factorization

Abstract

As data that have multiple views become widely available, the clusterization of such data based on nonnegative matrix factorization has been attracting greater attention. In the majority of studies, the statement in which all objects have images in all representations is considered. However, this is often not the case in practical problems. To resolve this issue, a novel semi-paired multiview clustering algorithm is proposed. For incomplete data, it is assumed that their views have the same indicator vector, and the paired matrix is introduced. The objects that are close to each other in each view must have identical indicators, which makes regularization and reconstruction of the manifold geometric structure possible. The proposed algorithm can work both with incomplete and complete data having multiple views. The experimental results obtained on four datasets prove its effectiveness compared to other modern algorithms.