Supervised Feature Discretization by Mutual Information Maximization
Figueiredo, M. A. T.
Supervised Feature Discretization by Mutual Information Maximization, Proc Portuguese Conf. on Pattern Recognition - RecPad, Coimbra, Portugal, Vol. --, pp. -- - --, October, 2012.
Digital Object Identifier:
Feature discretization (FD) techniques may be beneficial to many machine learning problems. FD leads to compact data representations, ignoring minor fluctuations that are irrelevant or even harmful for the learning task. Moreover, it is often the case that learning with discrete representations yields both lower training time and better accuracy. In this paper, we
propose a supervised FD technique based on the maximization of the mutual information (MI) between each discrete feature and the class label. The discretization intervals are obtained incrementally using a recursive procedure. Experimental results on standard benchmarks, with binary and multi-class problems, show that our method usually achieves better accuracy
than other well-known supervised FD approaches.