How it works
In unsupervised learning, an algorithm looks for structure in data without receiving a correct, pre-labeled answer for each example. Instead of learning to predict a known label, it tries to identify groups, relationships, compact representations, or observations that deviate from the pattern.
The term "unsupervised" describes the absence of target labels during training, not the absence of human decisions. People still choose the data, the variables, the algorithm, and the criteria to evaluate whether the result is useful. The scikit-learn documentation brings together techniques such as clustering, dimensionality reduction, and novelty detection.
In clustering, for example, the method groups records considered similar according to a chosen measure. In dimensionality reduction, it transforms data into a smaller representation that tries to preserve important aspects. The techniques have different assumptions and limitations: a cluster is a mathematical hypothesis about the data, not a true category discovered automatically.
The result can change with the scaling of variables, the handling of missing values, the choice of distance, and the method parameters. For that reason, preparing the data and testing the stability of findings is often as important as choosing the algorithm.
Practical example
An online store has purchase records but has not yet defined customer profiles. A team can remove unnecessary identifiers, select relevant features, and experiment with a clustering technique to observe patterns in frequency, product variety, and order value.
The method may group customers with similar habits. The team then needs to inspect the groups, check whether they make sense for the business, and talk to people who know the data. A group should not be automatically called "loyal customers" or "customers at risk": that meaning depends on evidence and a business decision.
To validate usefulness, you can check whether the groups repeat across different samples or periods and whether they help answer a concrete question, such as adapting the catalog organization. If the clustering is unstable or does not change any decision, there is no reason to treat it as a valuable discovery.
What it's for
This type of learning is useful when there are no reliable labels or when the initial goal is to explore a dataset. It can support exploratory segmentation, information compression, detection of unusual patterns, and analysis of documents or images.
It does not replace supervised learning for every task. If there is a known answer you want to predict, such as whether a payment was confirmed, the problem can be formulated as supervised classification. Sometimes a team starts without labels, examines patterns, and then creates human labels to train or evaluate another model.
Practical choice should start with the question, not the algorithm. Define what kind of structure would be useful, document the data transformations, compare more than one method when appropriate, and explain the uncertainty of the results. The output of an unsupervised model is an instrument for investigation, not a final decision about people or operations.