Data dependent evaluation of dissimilarities in nearest prototype vector quantizers regarding their discriminating abilities

Abstract

In this paper we propose a rank measure for comparison of (dis-)similarities regarding their behavior to reflect data dependencies. It is based on evaluation of dissimilarity ranks, which reflects the topological structure of the data in dependence of the dissimilarity measure. The introduced rank measure can be used to select dissimilarity measures in advance before cluster or classification learning algorithms are applied. Thus time consuming learning of models with different dissimilarities can be avoided.