Unsupervised Multi-View K-Means Clustering Algorithm

Abstract

Since advanced technologies via social media, internet, virtual communities and networks and internet of things (IoT), there are more multi-view data to be collected. Multi-view clustering is a substantial tool as a natural design for clustering multi-view data. K-means (KM) clustering for (single-view) data had been extended for handling multi-view data, called multi-view KM (MV-KM). In the literature, most MV-KM clustering algorithms are reported to be influenced by initializations and also need a given number of clusters. In this paper, we propose an unsupervised type of MV-KM clustering algorithm so that it can automatically find an optimal number of clusters without any initialization. We call it unsupervised MV-KM (U-MV-KM). Moreover, we also propose three multi-view cluster validity indices, called multi-view Dunn index (MV-Dunn), multi-view generalized Dunn index (MV-G-Dunn) and multi-view modified Dunn (MV-M-Dunn) indices for MV-KM clustering algorithms. We make experiments on some synthetic and real data sets and also make comparisons with some existing algorithms. Based on the experimental results and comparisons, the proposed U-MV-KM clustering algorithm actually shows good results. We also apply U-MV-KM to real data sets, the results demonstrate the superiority and usefulness of the U-MV-KM algorithm for clustering multi-view data.