Experiments on a representation-independent 'Top-Down and Preue' Induction Scheme

摘要

Recently,some methods for the induction of Decision trees have received much theoretical attention.While some of these works focused on efficient top-down induction algorithms,others investigated the pruning of large trees to obtain small and accurate formulae.This paper discusses the practical possibility of combining and generalizing both approaches,to use them on various classes of concept representations,not strictly restricted to decision trees or formulae built from decision trees.The algorithm,WIREi,is able to produce decision trees,decision lists,polynomials,and more.This shifting ability allows to reduce the risk of deviating from valuable concepts during the induction.As an example,in a previously used simulated noisy dataset,the algorithm managed to find systematically the target concept itself,when using an adequate concept representation.Further experiments on twenty-two readily available datasets show the ability of WIREi to build small and accurate concept representations,which lets the user choose his formalism to best suit his interpretation needs,in particular for mining purposes.