Supervised learning is a machine-learning paradigm in which a model is trained on labelled examples — inputs paired with known correct outputs — so that it can predict the output for new, unseen inputs. Its counterpart, unsupervised learning, works with unlabelled data and seeks to discover structure, patterns or groupings without being told what the “right” answer is. The presence or absence of labels is the defining distinction between the two.
Both belong to the wider field of machine learning, and choosing between them depends on whether you have labelled data and what question you are asking. A third paradigm, reinforcement learning, sits apart from both.
Supervised learning: learning from labels
In supervised learning, each training example carries a label: the email is “spam” or “not spam”, the image is “tumour” or “benign”, the house has a known sale price. The algorithm learns a mapping from inputs (features) to outputs (labels), and its performance is judged by how accurately it predicts labels for data it has not seen. Two main task types exist. Classification predicts a discrete category (spam or not, species A, B or C). Regression predicts a continuous quantity (a price, a temperature, a concentration).
Common supervised methods include linear and logistic regression, decision trees, support vector machines and neural networks. The key practical requirement — and often the key cost — is obtaining enough accurately labelled data, which may require expert annotation.
Unsupervised learning: finding structure
Unsupervised learning has no labels. Instead, the algorithm looks for inherent structure in the data. Clustering groups similar items together — for example, segmenting samples into subtypes without prior categories, using methods such as k-means or hierarchical clustering. Dimensionality reduction, such as principal component analysis, compresses many variables into fewer while preserving structure, aiding visualisation and downstream analysis. Because there is no ground-truth label, evaluating unsupervised results is harder and often relies on domain judgement.
Reinforcement learning: a third paradigm
Reinforcement learning differs from both. Here an agent learns by interacting with an environment, taking actions and receiving rewards or penalties, and gradually improving a policy to maximise cumulative reward. It is neither labelled-example learning nor pure pattern discovery; it learns from consequences over time. Reinforcement learning underlies advances in game-playing and robotics, and is noted here for completeness rather than treated in depth.
Comparing the paradigms
| Feature | Supervised learning | Unsupervised learning |
|---|---|---|
| Data | Labelled (input–output pairs) | Unlabelled |
| Goal | Predict known outputs | Discover hidden structure |
| Main tasks | Classification, regression | Clustering, dimensionality reduction |
| Example methods | Logistic regression, decision trees, SVMs | k-means, hierarchical clustering, PCA |
| Evaluation | Accuracy against held-out labels | Often qualitative; needs domain judgement |
Research uses and good practice
In research, supervised learning suits prediction and classification where labelled outcomes exist — diagnosing from images, predicting properties from features. Unsupervised learning suits exploration — finding subgroups, detecting anomalies, reducing dimensionality before further analysis. The two are often combined: unsupervised methods can pre-process or explore data that a supervised model then uses.
Whichever paradigm is used, the outputs are research outputs that require careful reporting: how data were labelled or collected, how the model was validated, and how results were evaluated. Sharing data, code and methods using consistent terminology supports reproducibility, and our guidance for authors covers documenting such computational work. For foundational background, see our overview of machine learning concepts and methods and the explainer on neural networks and deep learning.
Frequently asked questions
What is the main difference between supervised and unsupervised learning?
Labels. Supervised learning trains on labelled examples — inputs paired with known correct outputs — to make predictions, whereas unsupervised learning works with unlabelled data to discover structure such as clusters or lower-dimensional representations without any predefined answer.
What are classification and regression?
They are the two main supervised tasks. Classification predicts a discrete category, such as spam or not spam. Regression predicts a continuous value, such as a price or temperature. Both learn a mapping from input features to known output labels.
Where does reinforcement learning fit?
It is a separate, third paradigm. An agent learns by acting in an environment and receiving rewards or penalties, improving its policy over time to maximise cumulative reward. It learns from consequences rather than from labelled examples or pure pattern discovery.
Can the two approaches be combined?
Yes, frequently. Unsupervised methods such as clustering or PCA can explore or pre-process data, and a supervised model can then make predictions from the result. Many research pipelines use both, so they are complementary rather than mutually exclusive.