Data-Driven Constructive Induction

Constructive induction divides the problem of learning an inductive hypothesis into two intertwined searches: one—for the “best” representation space, and two—for the “best” hypothesis in that space. In data-driven constructive induction (DCI), a learning system searches for a better representation space by analyzing the input examples (data). The presented datadriven constructive induction method combines an AQ-type learning algorithm with two classes of representation space improvement operators: constructors, and destructors. The implemented system, AQ17-DCI, has been experimentally applied to a GNP prediction problem using a World Bank database. The results show that decision rules learned by AQ17-DCI outperformed the rules learned in the original representation space both in predictive accuracy and rule simplicity.

• Type: Journal
• Year: 1998
• Where: EXPERT
• Authors: Eric Bloedorn, Ryszard S. Michalski
• Keywords: Constructive Induction, Data-driven Constructive Induction, EXPERT 1998, Representation Space