An Ensemble Method for Clustering

Abstract

Combination strategies in classification are a popular way of overcominginstabilities in classification algorithms. A direct applicationof ideas such as “voting” to cluster analysis problems is not possible,as no a priori class information for the patterns is available. Wepresent a methodology for combining ensembles of partitions obtainedby clustering, discuss the properties of such combination strategiesand relate them to the task of assessing partition “aggrement”.