Flexible Multidiscretizer Based on Measures which are Used in Induction of Decision Trees
Cezary Ko mider · 2002
Discretization of continuous attributes offers a number of benefits for a machine learning process. Fundamental benefits are: a significant decreasing of learning time, a possible improvement of knowledge quality in case of noisy data, an increasing of knowledge legibility. In this paper we show analysis of a multidiscretizer which allows to apply a wide range of supervised discretization algorithms and which has very good abilities of tuning. Designed multidiscretizer allows to generate new supervised discretization algorithms and also variants similar to well known classic discretization algorithms such as e.g., ChiMerge. The multidiscretizer is based on top-down method and bottom-up one. It uses many measures which are mainly used in induction of decision trees and a few stop criterions. In this paper we study quality of the discretization generated by the multidiscretizer through evaluation of decision trees’ accuracy. We focus on comparison of top-down method with bottom-up one and also the measures for the configuration presented below. For discretization research we use decision tree induction algorithms C4.5 and ID3.