Semi-supervised learning

Semi-supervised learning is a paradigm in machine learning that trains a model on a small amount of labeled data together with a much larger amount of unlabeled data. Labels are often costly because they need a skilled person or a physical experiment, such as transcribing an audio segment or determining a protein structure, while unlabeled examples are cheap to collect, so combining the two can beat both supervised learning on the labeled part alone and unsupervised learning on the unlabeled part alone. The setting comes in two forms: transductive learning, which only tries to infer labels for the particular unlabeled examples at hand, and inductive learning, which tries to learn a general mapping from inputs to labels. Methods rely on assumptions that tie the shape of the data to the labels, namely that nearby points share a label, that data falls into clusters, or that it lies near a lower dimensional manifold. Self-training is the oldest approach, with applications from the 1960s, and Vladimir Vapnik formally introduced the transductive framework in the 1970s. Unlabeled data does not always help, because a poor match between the problem and the model assumption can lower accuracy.