A latent variable pairwise classification model of a clustering ensemble

Abstract

This paper addresses some theoretical properties of clustering ensembles. We consider the problem of cluster analysis from pattern recognition point of view. A latent variable pairwise classification model is proposed for studying the efficiency (in terms of "error probability") of the ensemble. The notions of stability, homogeneity and correlation between ensemble elements are introduced. An upper bound for misclassification probability is obtained. Numerical experiment confirms potential usefulness of the suggested ensemble characteristics. © 2011 Springer-Verlag.