Abstract

This paper presents a method for constructive induction in which new problem-relevant attributes are generated by analyzing consecutively created inductive hypotheses. The method starts by creating a set of rules from given examples using the AQ algorithm. These rules are then evaluated according to a rule quality criterion. Subsets of the best-performing rules for each decision class are selected to form new attributes. These new attributes are used to reformulate the training examples used in the previous step, and the whole inductive process repeats. This iterative process ends when the performance accuracy of the rules exceeds a predefined threshold In several experiments on learning different well-defined transformations, the method consistently outperformed (in terms of predictive accuracy) the AQ15 rule learning method, GREEDY3 and GROVE decision list learning methods. and REDWOOD and FRINGE decision tree learning methods.

Citations