Two-way induction

General-to-speci c learners like ID3 and CN2 perform well when the target concept descriptions are general, but often have di culties when they are speci c or mixed. This problem can be alleviated by combining them with a speci c-to-general learning component, resulting in a two-way induction system. In this paper one design for such a component is proposed, as well as two methods for combining the two components. Experiments on arti cial domains show the combined learner to match or outperform \pure" versions of C4.5 and CN2 across the entire generality spectrum, with the advantage increasing for greater concept speci city. Experiments on 24 real-world domains from the UCI repository con rm the utility of two-way induction: the combined learner achieves higher accuracy than C4.5 in 17 domains (at the 5% signi cance level in 12), and similar results are obtained with CN2. Closer observation of the system's behavior leads to a better understanding of its ability to correct overly-general rules with speci c ones, and shows that there is still room for improvement.