Cluster ensembles via weighted graph regularized nonnegative matrix factorization

Abstract

Cluster ensembles aim to generate a stable and robust consensus clustering by combining multiple different clustering results of a dataset. Multiple clusterings can be represented either by multiple co-association pairwise relations or cluster based features. Traditional clustering ensemble algorithms learn the consensus clustering using either of the two representations, but not both. In this paper, we propose to integrate the two representations in a unified framework by means of weighted graph regularized nonnegative matrix factorization. Such integration makes the two representations complementary to each other and thus outperforms both of them in clustering accuracy and stability. Extensive experimental results on a number of datasets further demonstrate this. © 2011 Springer-Verlag.