Discretization of Continuous Attributes for Learning Classification Rules

Abstract

We present a comparison of three entropy-based discretization methods in a context of learning classification rules. We compare the binary recursive discretization with a stopping criterion based on the Minimum Description Length Principle (MDLP)[3], a non-recursive method which simply chooses a number of cut-points with the highest entropy gains, and a non-recursive method that selects cut-points according to both information entropy and distribution of potential cut-points over the instance space. Our empirical results show that the third method gives the best predictive performance among the three methods tested.