Reliable instance classification with version spaces

Abstract

This paper proposes considering version spaces as an approach to reliable instance classification. The key idea is to construct version spaces containing the hypotheses of the target concept or its close approximations. So, the unanimous-voting classification rule of version spaces does not misclassify; i.e., instance classifications become reliable. We implement version spaces by testing them for collapse. We show that testing can be done by any learning algorithm and use support vector machines. The resulting combination is called version space support vector machines. Experiments show 100% accuracy and good coverage. © 2006 Springer-Verlag London.