A PARTITION TYPE METHOD FOR CLUSTERING MIXED DATA

Abstract

In this paper, we propose a method for clustering mixed data. The method is a nonhierarchical one, and deals simultaneously with variables of three main kinds: numerical, ordinal, and nominal. It is based on the minimization of a particular criterion f(G。) over all the partitions G。of n entities in m distinct clusters. The criterion is founded on a peculiar kind of internal standardized mean diversity of the entities, according to the three types of variables. The algorithm to get the best partition is also presented: it starts from a non-random choice of the first partition; the results are compared with those obtained by a random assignment to a first partition. In order to show the usefulness of the method and the performance of the algorithm on a large set of real data, an application to andrological mixed data is reported.