Unsupervised Learning

Unsupervised learning is a branch of machine learning where the model is given data but no labels, no correct answers, and no prior categories to sort things into. Instead of being told what to find, the model explores the data on its own, looking for natural groupings, relationships, or structures that are not immediately obvious to the human eye.

This approach is particularly valuable when you are dealing with large datasets and genuinely do not know what patterns might be hiding inside them. Rather than starting with a hypothesis, unsupervised learning lets the data suggest its own story.

How it works#

The two most common techniques in unsupervised learning are clustering and dimensionality reduction.

Clustering algorithms group data points that are similar to each other. The algorithm does not know in advance how many groups there should be or what they represent; it simply finds natural boundaries in the data. After the fact, a human analyst reviews the clusters and decides what they mean.

Dimensionality reduction is used when data has many features, too many to visualise or reason about directly. The technique compresses the data into fewer dimensions whilst keeping the most important structure intact. This can make patterns visible that were previously buried in noise, and it can also speed up other machine learning tasks by reducing the amount of data each model has to process.

A third technique, anomaly detection, uses unsupervised methods to identify data points that do not fit any cluster well. These outliers are often the most interesting cases: unusual transactions, rare equipment faults, or unexpected spikes in website traffic.

Examples#

Customer segmentation. A retail business holds years of purchase records but has no neat category system for its customers. An unsupervised clustering algorithm can analyse buying patterns across thousands of customers and group them automatically. The business might discover that one cluster shops exclusively online in the evenings, whilst another makes large in-store purchases at weekends. Each group can then be approached differently, without anyone having defined those segments in advance.

Fraud detection in banking. Banks feed transaction records into anomaly detection models. The model builds up a picture of what normal spending looks like for each customer. When a transaction falls far outside that pattern, it is flagged for review. Because fraudulent transactions are rare and varied, it would be impractical to label every possible fraud type in advance; unsupervised methods find the outliers without needing that prior knowledge.

Why it matters#

Unsupervised learning matters because most data in the world arrives without labels. Labelling is expensive and time-consuming, requiring domain expertise for every single example. Unsupervised methods can make sense of raw data at a scale that human labelling could never match.

It also serves as a foundation for more advanced AI systems. The pre-training phase of large language models is partly unsupervised: the model learns the structure of language from billions of words of text, with no human labelling individual sentences. Understanding this helps explain why these models can be so capable whilst also making errors that a labelled, supervised approach might have caught.

To see how unsupervised learning fits alongside other training approaches, visit AI vs ML vs Deep Learning, or browse all learning topics on the /glossary page.