---
title: 'Unsupervised learning'
description: 'Learning patterns, structure, or representations from data without supplied target labels.'
canonical_url: 'https://darkfactory.dev/glossary/unsupervised-learning'
markdown_url: 'https://darkfactory.dev/glossary/unsupervised-learning.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Unsupervised learning


## Definition

Unsupervised learning finds patterns, structure, distributions, or representations in data without a target label supplied for each example. Common tasks include clustering similar records, reducing dimensionality, estimating a probability distribution, detecting unusual cases, and learning latent factors.

The learning objective still encodes assumptions. A clustering algorithm may define similarity through Euclidean distance. A dimensionality-reduction method may preserve variance or local neighborhoods. An anomaly detector must decide what counts as unusual. The absence of labels does not make the result neutral or self-validating.

Evaluation is often harder than in supervised learning because there may be no single correct output. A useful evaluation can measure stability, reconstruction, likelihood, retrieval quality, separation against later labels, or performance on a downstream decision. A visually neat cluster is not evidence that it represents a meaningful customer, biological, or social category.

## Distinguish it from nearby terms

- **Supervised learning** fits examples paired with supplied targets.
- **Self-supervised learning** derives prediction targets from the data itself, such as masked or next tokens. It is often treated as part of the broader unlabeled-learning family, though current literature frequently names it separately.
- **Semi-supervised learning** combines labeled and unlabeled examples.
- **Clustering** is one unsupervised task. It is not a synonym for the whole field.
- **Exploratory data analysis** can reveal structure without training a model. Not every unlabeled analysis is machine learning.

## Operational significance

Name the structure the method is expected to recover and how it will be used. Check sensitivity to scaling, distance metric, initialization, random seed, and the number of groups or dimensions. Inspect whether missing data, high-cardinality fields, or protected attributes dominate the learned structure. If the output drives a decision about people, validate the categories with domain evidence rather than treating an algorithmic grouping as ground truth.

## Check your understanding

A clustering run divides customers into five groups. Before naming those groups as real customer types, test whether the assignments are stable, interpretable, and useful for a defined decision.

## Related terms

- [Supervised learning](https://darkfactory.dev/glossary/supervised-learning)
- [Self-supervised learning](https://darkfactory.dev/glossary/self-supervised-learning)
- [Clustering](https://darkfactory.dev/glossary/clustering)

## Evidence and further reading

- [NIST AI 100-2: Adversarial Machine Learning](https://csrc.nist.gov/pubs/ai/100/2/e2025/final)
- [NIST AI Resource Center Glossary](https://airc.nist.gov/glossary/)
- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)
- [Deep Learning](https://www.deeplearningbook.org/)