This paper proposes a new approach to classification reliability called version space
support vector machines (VSSVM). Their main advantage compared with previous
approaches to classification reliability is that they are not based on classificationreliability
parameter(s) and thus do not require learning any thresholds. The
experiments show that VSSVM achieve accuracy of 100% on the instances they
are able to classify when it is possible to maintain version spaces containing the
target hyperplane or its close approximations and the level of noise is low.
We foresee two future directions of research. The first one is to extend VSSVM
for classification tasks with more than two classes. The second direction is to apply
VSSVM when it is not possible to find consistent hyperplanes w.r.t. the training
data. In this context we consider to apply the generalized version spaces [7]