Using Neural Networks Within the Leaves of a Classification Tree

摘要

Classification trees and neural networks are widely used individually, yet little is known about the effect of combining these two techniques. Earlier work has shown that using k-nearest neighbor (k-NN) inside the leaves of a tree can increase classification accuracy. Since neural networks are so powerful, we apply neural networks instead of the k-NN method inside the leaves of the tree. This thesis studies the performance of this composite classifier. It is compared to the tree-structured classifier and the neural network classifier. We use commonly available data sets in this application and compare the results to those generated by other generally used classifiers. Compared to the results of the other two classifiers in this thesis, the composite classifier always gives the lowest cross-validated misclassification error rates in these data sets. Its excellent performance tells us that it is worth further investigation.