Balanced Neural Tree for Pattern Classification and Applications

Abstract

In comparison to standard Neural trees, this work presents a novel architecture of neural tree,” the balanced neural tree, for reducing tree size and enhancing classification. To obtain this result, there have been introduced two main innovations: (a) “perceptron substitution” (b) “pattern elimination”. The firstly introduced innovation attempts to bring the tree's structure into balance. If the supplied training set is significantly misclassified into a smaller no. of classes by the last-trained perceptron, it is replaced with a new-perceptron. The second innovation is the birth of a new criterion for the elimination of difficult training-patterns that cause over-fitting. Finally, to shorten perceptron training time, a new error-function based on tree depth is specified. On a variety of synthetic and real datasets, the suggested BNT has been evaluated. The results of the experiments reveal that the suggested BNT achieves satisfactory results in terms of tree depth reduction and classification accuracy.