What is semi-supervised learning?
Semi-supervised learning is machine learning that combines a small amount of labeled data with a large amount of unlabeled data. It sits between supervised learning, which needs a label on every example, and unsupervised learning, which uses none. The labeled examples anchor what the categories mean, and the unlabeled examples teach the model the shape of the data as a whole.
The most common approach is pseudo-labeling. A model is trained on the labeled set, then used to label the unlabeled records. Predictions it makes with high confidence are added to the training data and the model is retrained. Other methods train the model to give the same answer for an example and a slightly altered copy of it, which forces consistent behavior across the unlabeled data.
Semi-supervised learning in practice
A law firm wants to sort forty thousand archived agreements by type. Paralegals label six hundred of them. A classifier trained on those alone is mediocre, but after two rounds of pseudo-labeling across the full archive its accuracy on a held-out set rises to a usable level. The labeling budget stays at six hundred documents.
The method works best when examples of the same category cluster together in the data. When they do not, pseudo-labels can reinforce early mistakes, so keep a labeled holdout set to check each round. Active learning, which asks people to label the examples the model is least sure about, is a natural companion.