New asymmetric iterative scaling models for the generation of textual word maps

Abstract

The iterative spring model (Kopcsa and Schiebel, 1998) is a kind of multidimensional scaling algorithm (MDS) based on point mass mechanics, that embeds objects in a two dimensional Euclidean space and allows to visualize object relationships and cluster structure. This technique assumes that the similarity matrix for the data set under consideration is symmetric. However there are many interesting problems where asymmetric proximities arise, like text mining problems. In this work we propose a variety of improvements to this algorithm to deal with asymmetric dissimilarities. Clustering quality and distances preservation of the resulting word maps are evaluated through objective measures. The new asymmetric algorithms outperform both, their symmetric counterpart and other widely used multidimensional scaling methods according to the objective measures computed.