Variable selection in cluster analysis: an approach based on a new index

In cluster analysis, the inclusion of unnecessaryvariables may mask the true group structure. For the selection ofthe best subset of variables, we suggest the use of two overallindices. The first index is a distance between two hierarchicalclusterings and the second one is a similarity index obtained asthe complement to one of the previous distance. Both criteria canbe used for measuring the similarity between clusterings obtainedwith different subsets of variables. An application with a realdata set regarding the economic welfare of the Italian Regionsshows the benefits gained with the suggested procedure.