Glossary
What is unsupervised learning?
In short
Unsupervised learning is a way of training in which a model finds groups and patterns in data on its own, without answers labeled in advance.
AI basics · unsupervised learning
MMOOZZGG.com editorsChecked: 1 min read
For a child
Pour a box of different buttons onto the carpet and ask a friend to sort them into piles. You do not say how. Your friend decides: big with big, red with red, the ones with four holes on their own. Nobody gave the right answer; your friend just noticed how the buttons are alike. A computer can sort things that nobody labeled in the same way.
In plain words
Sometimes you have a lot of data but no labels: just thousands of buyers, calls or photos. Then the model is not trained to guess an answer. It is asked to bring order to the pile. It compares objects by their features and puts similar ones into groups, notices odd records that look like nothing else, or shrinks a hundred table columns to the two or three that matter most. This way a food delivery service finds out, with no hints at all, that its customers include night-owl students, families with children and offices ordering lunch. A person names the groups afterwards.
For business
The method helps when there are no labels or labeling is expensive. Companies use it to split customers into segments for marketing, find suspicious payments and equipment faults, group support requests by topic and find duplicates in product catalogs. Most of the cost goes to data preparation and to an analyst who checks whether the groups make business sense. The risk is that the algorithm always finds something, even in random noise, and nice segments bring no money. So compare the result with sales on a small sample before you change processes. A pilot is a convenient format for this check.
For a specialist
Unsupervised learning estimates the structure of a data distribution without a target variable. Its main tasks are clustering, dimensionality reduction, density estimation and outlier detection. The scikit-learn user guide lists k-means, DBSCAN and hierarchical clustering, Gaussian mixture models, principal component analysis (PCA), non-negative matrix factorization (NMF) and manifold learning. For visualizing high-dimensional data, t-SNE is common: van der Maaten and Hinton described it in the Journal of Machine Learning Research in 2008. A close field, self-supervised learning, builds the task from the data itself, for example predicting a hidden word; the pretraining of language models and embeddings relies on it. Measuring quality is harder here: internal metrics such as the silhouette score and expert review are used.
Example
A chain of fitness clubs loads visit data without any labels. The algorithm finds a group of members who used to come three times a week but came only once a week in the last month. The club offers them a personal training session before they quit.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map