EXPLAINER

Unsupervised Learning

conceptfreshness: stablesources not yet verified

Unsupervised learning finds structure in unlabelled data — clusters, patterns, compact representations — without being told the right answers.

Short answer: Unsupervised learning looks for structure without labels: clustering, dimensionality reduction, density estimation, anomaly detection. Evaluation is harder because there is no single correct answer.

What is it?

With no answer key, the model looks for patterns itself: groups of similar customers, unusual transactions, or a simpler way to represent complex data. Results need human interpretation — there's no label to score against.

Also known as: clustering, k-means, dimensionality reduction, PCA, anomaly detection, unlabelled data learning, unsupervised machine learning.

Technical definition

Tasks include clustering (k-means, hierarchical, DBSCAN, Gaussian mixtures), dimensionality reduction/visualisation (PCA, t-SNE, UMAP), density estimation and anomaly detection, and representation learning with autoencoders. Evaluation relies on internal criteria (silhouette, reconstruction error), downstream task performance, or expert review; different algorithms embed different assumptions about cluster shape and scale.

Entity type: concept · Freshness class: stable.

When should I use it?

Use for exploration, segmentation, compression and anomaly detection when labels are unavailable.

Implementation patterns

  • Scale features, then cluster.
  • Reduce with PCA before visualising with t-SNE/UMAP.
  • Cluster embeddings of text or images to explore a corpus.
  • Anomaly scoring with isolation forests or reconstruction error.

Technologies and frameworks

Commonly associated: scikit-learn, UMAP, HDBSCAN, PyTorch autoencoders. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Clarify the question the structure should answer.
  2. Clean and scale features.
  3. Try more than one algorithm.
  4. Inspect results by sampling.
  5. Validate usefulness with a downstream task or expert.

Common failure modes

  • Unscaled features dominating distances.
  • Treating a visualisation as proof of real clusters.
  • Choosing k arbitrarily.
  • Clusters that are statistically fine but not useful.

Troubleshooting

  • Clusters look arbitrary — Scale features, try other algorithms, inspect samples per cluster.
  • t-SNE plot changes each run — Fix seeds and treat distances in it as non-metric.

Contrasted with

Questions this page answers

  • What is unsupervised learning?
  • How does k-means clustering work?
  • What is PCA used for?
  • How do I choose the number of clusters?

Typical goals: group similar items without labels; reduce data dimensions; detect unusual records.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: stable — Core ideas that change slowly (maths, classic algorithms). Re-review yearly.

Canonical sources (unverified leads):

  • Hastie et al., The Elements of Statistical Learning, ch. 14 — hastie.su.domains/ElemStatLearn
  • scikit-learn clustering and decomposition guides — scikit-learn.org