Training data is used for machine learning. In the case of supervised learning the training data conisists of labelled input–output pairs, which the machine learning attempts to emulate and generalise. In the case of unsupervised learning the training data contains inputs only without any form of labelling so that the machine learning is looking for patterns and similarities in the data.
In real life, available training data may be partially labelled, with some input–output pairs and some plain inputs. Furthermore there may be missing values amongst the input features.
Training data acts as a sample of the expected set of all possible inputs, but this may be hard to achieve in practice. Sometimes the nature of data collection methods may mean that some groups are underrepresented, leading to potential bias. A totally random sample of inputs can also be problematic as infrequent classes may not appear at all, for example, people with a rare disease; to avoid this these minority or rare groups may be deliberately over-represented in the training data, with some form of weighting or correction applied during or after training.
Used in glossary entries: bias, labelling, machine learning, supervised learning, unsupervised learning