Auto-associative multivariate regression trees for cluster analysis

Abstract

Multivariate Regression Trees, an intuitive and simple regression technique, intrinsically produce homogenous subsets of data. These characteristics imply that Multivariate Regression Trees have the potential to be utilised as an easily interpretable clustering method. The suitability of Multivariate Regression Trees as a clustering technique is investigated with two real datasets containing only explanatory variables. The preliminary results show that Multivariate Regression Trees as a clustering algorithm produce clusters of similar quality to the well-known K-means technique, and more recent approaches to Cluster Analysis including Mixture Models of Factor Analysers and Plaid Models. The study also evaluates the suitability of various criteria used to describe cluster solutions.