unsupervised learning

basicwidespread

A type of machine learning where the program, without pre-defined correct answers, independently searches for groups, patterns, and outliers in the data.

In plain terms

Imagine receiving a large box of mixed LEGO bricks without instructions. No one tells you what is what, yet you can sort them by color, size, or shape. Unsupervised learning does exactly this with data: it receives no answer key, only the raw material, and seeks order within it.

It is useful for spotting connections you did not anticipate: which shoppers resemble each other, which products are bought together, and which transaction deviates from the norm. It is most useful where there is vast data but no one has labeled it.

In supervised learning, the machine receives examples along with the correct answer, such as 'this image is a cat, that one is a dog.' There is no such teacher correction here. Therefore, the result cannot simply be ticked off: the machine provides clusters, but a human must judge what they mean and whether they are useful.

How it works

Data is collected without labels and converted into numbers, for example, describing a customer by their age, spending, and purchase frequency.
The algorithm calculates which data points resemble each other, meaning how close they are in this number-based space.
It searches for structure based on similarity: it organizes data into groups (clustering), compresses them into fewer features (dimensionality reduction), or flags cases that deviate significantly from the rest (anomaly detection).
Many methods, such as k-means, recalculate the partition multiple times until the groups change little; other procedures yield results in a single calculation.
At the end, a human reviews the results, names the groups, and decides whether they have practical value.

An everyday example

When the online store displays "people who bought this also bought this," or the music app recommends favorites of listeners with tastes similar to yours, such a method often works in the background. No one manually assigned you to a group: the system found out from your behavior who you resemble.

Who is affected

marketerCustomer segmentation typically works like this: the algorithm proposes customer groups from the data rather than starting from your assumptions. But you have to check whether the groups make real sense. → Marketing

bank customerBanks detect unusual, fraud-suspect transactions partly using this method, so a strange purchase of yours might be stopped.

industrial worker, engineerThis is how unusual behavior can be spotted in machines' sensor data, even before a failure occurs. → Industry

developerIf you lack the funds or time to label data, this is the first tool to help you understand the dataset's structure. → Code

A common misconception

MythUnsupervised learning means the machine learns and decides independently without human intervention.

In realityHere 'supervision' simply means that no pre-defined correct answer exists for the data. A human still decides what data the machine receives, which method it uses, and what the result means.

Limits and risks

There is no single clear good solution, making it difficult to measure whether the resulting grouping is truly correct or just appealing.
The algorithm finds clusters even if there is no meaningful structure in the data, so the pattern is not proof.
The result depends heavily on the settings, for example, how many groups you request and which features you specify.
If the data is biased, the machine learns the bias as a pattern, and in classifications affecting people, this can lead to discriminatory treatment.

Similar, but not the same

supervised learning – There, every training example has a correct answer the machine aligns with; here, there is no answer key.
self-supervised learning – In self-supervised learning, the machine generates tasks and correct answers from the data itself (for example, guessing the next word), whereas classic unsupervised methods only search for structure.
reinforcement learning – There, the machine tries and receives a reward or penalty for its steps; here, there is no feedback at all.

Related concepts

Broader

machine learning, artificial intelligence

Narrower

clustering, dimensionality reduction, anomaly detection

On the edge

supervised learning, self-supervised learning, reinforcement learning, training data, data labeling

A brief history

The roots of the method family go back to statistics: Karl Pearson described principal component analysis in 1901, which remains one of its basic tools today. The best-known clustering method, k-means, takes its name from James MacQueen's 1967 study.

Where things stand

As of late 2026, it is a common, everyday tool in recommendation systems, fraud detection and data analysis. Large language models are also pre-trained on unlabeled data, which is now mostly called self-supervised learning.

Status: 2026 November

Also known as

unsupervised learning · unsupervised learning · unsupervised learning · unsupervised machine learning · unsupervised learning · unsupervised learning · unsupervised machine learning · unsupervised ML

Sources

scikit-learn documentation: Unsupervised learning

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.