Definition
Supervised learning trains a model on examples that pair inputs with target outputs. A spam filter may learn from messages labeled spam or not spam. A demand model may learn from product and calendar features paired with observed sales. Training adjusts the model to reduce a loss between its predictions and those targets.
Classification and regression are the most common supervised task families. Classification targets categories. Regression targets numeric values or distributions. Structured prediction extends the setup to sequences, trees, or other outputs such as named entities in a sentence.
The labels can come from people, measurements, business records, rules, or another system. "Supervised" describes the training signal, not the label's quality. Historical outcomes may encode inconsistent judgment, missing cases, policy changes, or social bias. A label can also be a proxy that differs from the decision the organization actually cares about.
Distinguish it from nearby terms
- Self-supervised learning derives targets from the data's own structure, such as the next or masked token.
- Unsupervised learning looks for structure without target labels for each example.
- Semi-supervised learning combines labeled and unlabeled examples.
- Reinforcement learning optimizes sequential behavior from rewards and transitions rather than fixed input-output pairs.
- Human feedback can produce supervised preference labels, demonstrations, or reward-model data. The presence of feedback does not identify the training method by itself.
Operational significance
Keep training, validation, and test examples separate so evaluation measures generalization rather than memory. Define who or what produced each label, how disagreements were handled, and whether the label is available at prediction time. Check subgroup performance and class balance. When production conditions or policies change, labels from the past may no longer represent the target relationship.
Check your understanding
A model predicts whether invoices need manual review. If reviewers' past decisions are used as labels, the model learns those decisions, including any inconsistency in them. It does not automatically learn the true risk of each invoice.