Transductive learning

When a semi-supervised model is aimed at finding the labels for the unlabeled samples, the approach is called transductive learning. In this case, we are not interested in modeling the whole distribution p(x|y), which implies determining the density of both datasets, but rather in finding p(y|x) only for the unlabeled points. In many cases, this strategy can be time-saving and it's always preferable when our goal is more oriented at improving our knowledge about the unlabeled dataset.