ON DETECTING CLUSTERS USING PROJECTION PURSUIT METHODS

Abstract

Projection pursuit methods are aimed at searching for non-linear structure through linear projections from a p dimensional space into a subspace of dimension q, usually the plane or line. In the context of Cluster analysis, these methods can be used to detect grouped structure; though not all non-linear structures correspond to grouped structure. There are various methods of Projection pursuit, each one is tied to an index of projection, and most indices are measures of departures from normality. In this paper we present various indices of projection, referring to existing literature for details and discussion on their theoretical properties, and illustrate their use with an example from a study of regeneration of native vegetation in the Amazonian region of Brazil. Here the problem was to identify homogeneous groups of plant families; four groups were suggested and identified.