Experiments with Decision Tree Classifiers – Discretization of Numerical Attributes

Classification algorithms are used in numerous applications everyday, from assigning letter grades to student student’s scores, to computerized letter recognition in mail processing. Discretization consists of applying a set of rules to reduce the number of discrete intervals from which an attribute is assigned. Discretization is generally applied to datasets whose numerical range consists of continuous values, replacing the numerical domain with a nominal domain. This paper’s primary focus is on the application of discretization algorithms on Classification and Regression Trees (CART) using Equal Width, Equal Frequency and CAIM discretization algorithms. Wile no method dominated others for all datasets, results show that discretization prior to classification generally decreases the processing time, while increasing the accuracy of the classifier. The data presented here further substantiates this claim. Index Terms — Supervised discretization, CART, classification, class-attribute interdependency, continuous attributes, machine learning algorithm.