Foundations
Unsupervised Learning
Also known as: learning without labels
Finding patterns and structure in data that has no labels or correct answers attached.
Unsupervised learning is when a computer finds patterns on its own, without being told the right answers. Give it a pile of customer data and it might discover natural groups of similar customers that no one had defined in advance. There is no answer key here; the model's job is simply to find structure hidden in the data.
Unsupervised learning works with unlabelled data. Instead of predicting a known target, the model discovers structure, groupings, relationships or a simpler representation, by itself. It contrasts with supervised learning, which needs labelled examples.
Common tasks
- Clustering: group similar items (customer segments, topics in documents).
- Dimensionality reduction: compress data to its most important features for visualisation or efficiency.
- Anomaly detection: flag items that do not fit the usual pattern (fraud, faults).
Why it is valuable
Unlabelled data is abundant and cheap; labelled data is scarce and costly. Learning useful structure without labels is therefore powerful, and it underpins embeddings, the meaning-vectors behind semantic search and RAG.
Unsupervised learning models the structure of unlabelled data, estimating densities, latent factors or groupings without target supervision. Canonical methods include clustering (k-means, hierarchical, DBSCAN, Gaussian mixtures), dimensionality reduction (PCA, t-SNE, UMAP, autoencoders), and generative density estimation. Evaluation is intrinsically harder than in supervised settings because there is no ground-truth label to score against.
The modern boundary with unsupervised learning is self-supervised learning, which constructs a supervised objective from the data itself (masking, next-token prediction, contrastive pairing). This is the engine of foundation-model pretraining: it turns raw, unlabelled corpora into a scalable training signal, which is why LLMs could be trained on internet-scale text without human annotation. Learned representations, embeddings, are a chief practical output, reused across downstream tasks.