A quality-driven ensemble approach to automatic model selection in clustering

Abstract

A fundamental limitation of the data clustering task is that it has an inherent, ill-defined model selection problem: the choice of a clustering technique also implies some a-priori decision on cluster geometry. In this work we explore the combined use of two different clustering paradigms and their combination by means of an ensemble technique. Mixing coefficients are computed on the basis of partition quality, so that the ensemble is automatically tuned so as to give more weight to the best-performing (in terms of the selected quality indices) clustering method. © Springer International Publishing Switzerland 2014.