Comparison of Discretization Methods for Preprocessing Data for Pyramidal Growing Network Classification Method

Abstract

This paper presents a comparison of four representative discretization methods from different classes to be used with so called PGN-classifier which deals with categorical data. We examine which of them supplies more convenient discretization for PGN Classification Method. The experiment are provided on the base of UCI repository data sets. The comparison tests were provided using an experimental classification machine learning system "PaGaNe", which realizes Pyramidical Growing Network (PGN) Classification Algorithm. It is found that in general, PGN-classifier trained on data preprocessed by Chi-merge achive lower classification error than those trained on data preprocessed by the other discretization methods. The comparison of PGN-classifier, trained with Chi-merge discretizator with other classifiers (realized in WEKA system) shows good results in favor of PGN-classifier.