Using a small amount of labelled data together with a large amount of unlabelled data. It suits the very common situation where collecting examples is cheap and having experts label them is expensive. Medical imaging is the classic case: plenty of scans, very few annotated by a consultant.
Technical view
Relies on assumptions linking the unlabelled distribution to the decision boundary, typically the cluster or manifold assumption. Methods include self-training, consistency regularisation under augmentation, and graph-based label propagation. Gains are real when the assumptions hold and negative when they do not, since confidently propagating wrong labels degrades performance below the supervised baseline.
Related
Part of AI Terminology Index, section: Classical Machine Learning.