A relevancy filter for constructive induction

Nada Lavrac, Dragan Gamberger, and Peter Turney. IEEE Intelligent Systems, 13(2):50--56, March-April 1998.

Abstract

Some machine learning algorithms enable the learner to extend its vocabulary with new terms if, for a given a set of training examples, the learner's vocabulary is too restricted for solving the learning task. In this work we propose a filter that selects the potentially relevant terms from the set of constructed terms, and eliminates the terms which are irrelevant for the learning task. By biasing constructive induction (or predicate invention) to relevant terms only, the explored space of constructed terms can be much larger. The elimination of irrelevant terms is specially well suited for learners of large time or space complexity (such as genetic algorithms and neural nets). This paper presents the REDUCE algorithm for eliminating irrelevant terms and a case study in which REDUCE is used to preprocess data for a hybrid genetic algorithm RL-ICET.

BibTeX entry.

Other publications

N Lavrac, Nada.Lavrac@ijs.si,