Quality evaluation of a selected partition : An approach based on resampling methods

Abstract

The aim of this work on cluster analysis is to provide a methodology to analyse and assess the quality of a selected partition (the best partition according to several validation indexes). In the proposed approach, the evaluation of the stability and of the consistency of the results of the selected partition (original partition) was done using the comparison between this partition and each of the partitions (with the same number of clusters that the original one) obtained by resampling. A special emphasis is given to an index defined by linear combination of four indicators, which allows evaluating the adjustment between the original partition and each of the partitions (and / or set of obtained partitions) obtained from resampling data. The application of these indexes is exemplified using a set of real data, and the main conclusions are summarized and discussed.