A Tuning Aid for Discretization in Rule Induction

This paper examines where a tuning aid can be useful to help discretization of numerical attributes in rule induction, and subsequently improve deduction of induction results. Diierent discretizationmethods use diierent strategies to set up the borders for continuous attributes. They mostly incorporate class supervision to deene the discretization borders. The tuning aid we present uses an unsupervised method to deene the intervals at induction time. We then supervise the learning process, by comparing the performance in terms of predictive accuracy with the information gain based discretization methods implemented in HCV (Version 2.0) and C4.5.