The idea behind semi-supervised learning is to learn from labeled and unlabeled data to improve the predictive power of the models. The notion is explained with a simple illustration, Figure 1, which shows that when a large amount of unlabeled data is available, for example, HTML documents on the web, the expert can classify a few of them into known categories such as sports, news, entertainment, and so on. This small set of labeled data together with the large unlabeled dataset can then be used by semi-supervised learning techniques to learn models. Thus, using the knowledge of both labeled and unlabeled data, the model can classify unseen documents in the future. In contrast, supervised learning uses labeled data only:
Figure 1. Semi-Supervised Learning process (bottom) contrasted with Supervised Learning (top) using classification of web documents as an example. The main difference is the amount of labeled data available for learning, highlighted by the qualifier...