Multi-view clustering by virtually passing mutually supervised smooth messages

Abstract

While the existing multi-view affinity propagation (AP)-based clustering method inevitably works with more than one random initialization and parameter, a novel algorithm called MVCPMM is proposed from a new perspective to achieve more consistent multi-view clustering results with only one random initialization and one parameter. The proposed virtual function nodes added between the variable nodes of the subgraphs of the AP factor graphs (i.e., individual views), enable the core of MVCPMM to pass mutually supervised smooth messages across different views and subsequently exchange the messages within individual views in a mutually supervised manner to encourage the clustering quality of individual views. In addition to maintaining the cluster diversity of individual views, MVCPMM penalizes the changes in the cluster structures of different views by using mutually supervised smooth messages as bidirectional cross-view messages, which can effectively improve the consensus of exemplars across different views. Experimental results on both synthetic and benchmark multi-view datasets demonstrate the superiority of MVCPMM in contrast to several state-of-the-art multi-view clustering methods in terms of both clustering performance and clustering consistency across different views.