Unsupervised Learning
Finding structure in data nobody labelled — useful, underrated, and much harder to know if you got right.
Reviewed July 10, 2026Stable
When not to use it
- When you have labels. If you know the categories, supervised learning is more accurate and you can actually measure it.
- When you need a defensible answer. "The algorithm found these groups" is not a justification a regulator or a board will accept without a domain expert vouching for them.
- Expecting the algorithm to tell you what matters. It finds structure under its own assumptions. You supply the meaning.
Reach for something else instead
- Supervised learning if labelling even a few hundred examples is feasible. It usually is, and it's usually worth it.
- Self-supervised methods when you want representations from unlabelled data — this is what actually works at scale now.
- Just looking at the data. Sorting, cross-tabbing, and plotting answers more questions than people expect, and it's honest about what it found.
This entry is part of a longer guide: What is machine learning?
Read more on the blog
- Supervised vs unsupervised learning: the four typesMachine learning is usually taught as three types: supervised, unsupervised, and reinforcement learning. That map is still useful, but it no longer covers the paradigm that trains almost every modern AI system. Understanding what the fourth type is, and why it broke the old split, explains how AI actually got here.
- What is reinforcement learning? Learning from rewardReinforcement learning went from a niche corner of AI obsessed with games and robots to the paradigm that shapes how every modern language model behaves. Here's what it actually is, learning by trial, reward, and consequence, why it's different from other machine learning, and how it quietly became the layer between a smart model and a useful one.
Further reading
- Hastie, Tibshirani & Friedman, The Elements of Statistical Learning, ch. 14 — the reference treatment, free from the authors.
- Wattenberg, Viégas & Johnson (2016), How to Use t-SNE Effectively — why the plots mislead, shown interactively.
- von Luxburg, Williamson & Guyon (2012), Clustering: Science or Art? — the evaluation problem, stated honestly.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Not scaling features, then discovering the clusters are entirely about whichever column had the biggest numbers.
- Reading distances and cluster sizes off a t-SNE or UMAP plot. Neither is meaningful; the plot is a projection, not a map.
- Accepting the clusters because the silhouette score was good. That measures agreement with the algorithm's assumptions, not truth.
At a glance
FieldMachine Learning
Core ideastructure without labels
Usessegmentation, anomalies, compression
Hard partknowing if it's real
DifficultyIntermediate
Often compared with
Unsupervised vs. supervised — asking what's in the data vs. teaching it what you already know.
Where this sits
1 concept come first. Understanding it opens up 89 more.
1Levelsteps in
1Needs firstconcepts
89Opens up29% of 310
11Areascrosses fields
Learn these firstTraining Data
Computed from the prerequisite graph, not assigned. How this works