Learning from a Test Set

Classication of partially labeled data requires linking the unlabeled input distribution P(x) with the conditional distribution P(y|x) obtained from the labeled data. The latter should, for example, vary little in high density regions. The key problem is to articulate a general principle behind this and other such reasonable assumptions. In this paper we provide a new approach to semi-supervised learning based on the stability of estimated labels for the unlabeled dataset, e.g a large test set, and the maximization of the mutual label relation. No clustering assumptions are required and the approach remains tractable even for continuous marginal class densities. We demonstrate the approach on synthetic examples and UCI repository datasets.