Semiüberwachtes

Semi-supervised learning, or semiüberwachtes Lernen, is a machine learning paradigm that combines a small set of labeled data with a larger pool of unlabeled data to build predictive models. It assumes that unlabeled data contain information about the structure of the input space that can improve generalization beyond what labeled data alone yield.

Assumptions commonly guiding semi-supervised methods include the smoothness or manifold assumption, the cluster assumption (points in the same cluster or on the same manifold are likely to share labels), and the low-density separation assumption (decision boundaries should lie in regions with few unlabeled points). Based on these, approaches include self-training, where a model iteratively labels unlabeled examples; co-training, where multiple views of the data teach each other; graph-based methods such as label propagation and label spreading that spread labels through a similarity graph; and semi-supervised versions of generative models and support vector machines. In deep learning, techniques like pseudo-labeling and consistency-based training are widely used, and graph neural networks extend semi-supervised learning to structured data.

Applications span natural language processing, computer vision, speech recognition, and bioinformatics, especially when labeled data are scarce or costly to obtain. Performance gains depend on the validity of assumptions and the distribution of data; poorly chosen unlabeled data can degrade accuracy.

Challenges include the risk of error amplification from incorrect labels, domain mismatch between labeled and unlabeled data, model bias, and computational complexity. The field has evolved from early methods in the 1990s to contemporary deep-learning techniques, with ongoing research into robust and scalable algorithms.