Abstract
This paper proposes a new approach to classification reliability. The key
idea is to maintain version spaces containing (close approximations of) the
target classifiers. In this way the unanimous-voting rule applied on these
version spaces outputs reliable instance classifications.
Version spaces are defined in a hypothesis space of oriented hyperplanes.
The unanimous-voting rule is implemented by testing whether version spaces
are empty. Testing is done by support vector machines. Hence, the approach
is called version space support vector machines.
Experiments with the approach show a 100% accuracy on the classified
instances at the cost of not classifying all instances.