Analysing Hierarchical Data Using a Stochastic Evolutionary Neural Tree

Abstract

SCENT is simple competitive neural network model that evolves a tree structured set of nodes in response to being presented with an unlabelled data set. The resulting set of weight vectors and their relationship can be viewed as giving a hierarchical classification of the training data. This paper examines the nature of this classification for two data sets over several runs of the network. The first data set is a set of grey scale images, chosen because the code-vectors produced by SCENT can then be visualised in a natural way. The second data set is a small set of vectors coding attributes of animals. The resulting taxonomy from SCENT can then be compared with the normal taxonomic groups that such a set of animals would fall into. Since the SCENT model is stochastic different runs produce different trees, but the variation in results produced over several runs is small. The model is shown to be reasonably robust and the relationship between the nature of the data and the type of tree produced is examined.