A discretization algorithm based on a heterogeneity criterion

Discretization, as a preprocessing step for data mining, is a process of converting the continuous attributes of a data set into discrete ones so that they can be treated as the nominal features by machine learning algorithms. Those various discretization methods, that use entropy-based criteria, form a large class of algorithm. However, as a measure of class homogeneity, entropy cannot always accurately reflect the degree of class homogeneity of an interval. Therefore, in this paper, we propose a new measure of class heterogeneity of intervals from the viewpoint of class probability itself. Based on the definition of heterogeneity, we present a new criterion to evaluate a discretization scheme and analyze its property theoretically. Also, a heuristic method is proposed to find the approximate optimal discretization scheme. Finally, our method is compared, in terms of predictive error rate and tree size, with Ent-MDLC, a representative entropy-based discretization method well-known for its good performance. Our method is shown to produce better results than those of Ent-MDLC, although the improvement is not significant. It can be a good alternative to entropy-based discretization methods.