Optimising a Neural Tree Using Subtree Retraining

Abstract

Subtree retraining applied to a Stochastic Competitive Evolutionary Neural Tree model (SCENT) is introduced. This subtree retraining process is designed to improve the performance of the original model which provides a hierarchical classification of unlabelled data. The effect of subtree retraining on the network produces stable classificatory structures by repeatedly restructuring the weakest branch of the classification tree based on internal relation between members. An experimental comparison using well-known real world data sets, chosen to provide a variety of clustering scenarios, showed the new approach produced more reliable performances.