Version Space Support Vector Machines

The task of reliable classification is to determine if a particular instance classification is reliable. There exist two approaches to the task: the Bayesian framework [3] and the typicalness framework [5]. Although both frameworks are useful, the Bayesian framework can be misleading and the typicalness framework is classifier dependent. To overcome these problems we argue to use version spaces (VSs) [3] for reliable classification. The key idea is to construct VSs containing hypotheses of the target class or of its close approximations. In this way the unanimous-voting classification rule of VSs does not misclassify instances; i.e., instance classifications become reliable. To construct VSs with the property above we propose an approach extending the volumes of VSs s.t. instance misclassifications are blocked. The approach and VSs are realized using support vector machines (SVMs) [4]. The combination is called version space support vector machines (VSSVMs). We show that VSSVMs are able to outperform the existing approaches to reliable classification.