Average Cluster Consistency for Cluster Ensemble Selection

Various approaches to produce cluster ensembles and several consensus functions to combine data partitions have been proposed in order to obtain a more robust partition of the data. However, the existence of many approaches leads to another problem which consists in knowing which of these approaches to produce the cluster ensembles' data and to combine these partitions best fits a given data set. In this paper, we propose a new measure to select the best consensus data partition, among a variety of consensus partitions, based on the concept of average cluster consistency between each data partition that belongs to the cluster ensemble and a given consensus partition. The experimental results obtained by comparing this measure with other measures for cluster ensemble selection in 9 data sets, showed that the partitions selected by our measure generally were of superior quality in comparison with the consensus partitions selected by other measures.