Hybrid trees for supervised learning : towards extraction of decision rulesFlorence d ' Alch

We propose a new constructive approach, hybrid trees, that combines decision trees and neural learning methods to learn and build automatically neural systems for classiication and regression tasks. Neural trees can be described as decision trees where each node is a formal neuron. In hybrid trees, internal nodes segment the input space into regions where the terminal nodes perform the classiication or regression task. Diierent kinds of functions can be used as terminal nodes such formal neurons, radial basis functions, small multilayered networks. As in simple decision trees, once the tree is built, one can extract decision rules associated with each leaf of the tree. The interpretability of the rule depends on the choice of the terminal nodes. We describe the hybrid trees in the case of a binary classiication task. As an example, application to handwritten character discrimination is presented.