Rule Induction and Instance-Based Learning: A Unified Approach

Abstract

This paper presents a new approach to inductive learning that combines aspects of instancebased learning and rule induction in a single simple algorithm. The RISE system searches for rules in a specific-to-general fashion, starting with one rule per training example, and avoids some of the difficulties of separate-andconquer approaches by evaluating each proposed induction step globally, i.e., through an efficient procedure that is equivalent to checking the accuracy of the rule set as a whole on every training example. Classification is performed using a best-match strategy, and reduces to nearest-neighbor if all generalizations of instances were rejected. An extensive empirical study shows that RISE consistently achieves higher accuracies than state-of-the-art representatives of its "parent" paradigms {PEBLS and CN2), and also outperforms a decision-tree learner {C4.5) in 13 out of 15 test domains {in 10 with 95% confidence).