Unsupervised Learning
Finding structure in data that has no labels: groups, compressed representations and unusual points.
Someone dumps a giant box of LEGO on the floor with no instructions. Without being told, you start sorting it into piles: wheels here, windows there, small red bricks over there.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Techniques that apply here
Used by
How it works
Without labels there is nothing to predict, so the goal shifts to describing the data:
- Clustering: which points belong together? k-means is the classic.
- Dimensionality reduction: can 100 columns be summarized by 3 without losing much? PCA and autoencoders do this.
- Anomaly detection: which points do not fit any pattern?
Unsupervised learning is often a first exploratory step, or a way to create useful features for a later supervised model. Its close relative, self-supervised learning, also uses unlabeled data but invents its own prediction tasks, and it is how modern foundation models are trained.
Everyday use cases
- Case 01
Customer segmentation
Group shoppers by behavior to discover types like "bargain hunters" and "weekend bulk buyers".
- Case 02
Anomaly detection
Flag server logs or bank transactions that look unlike anything seen before.
Strengths, weaknesses and when to use it
Strengths
- No labeling cost; raw data is usually plentiful.
- Can reveal structure nobody thought to look for.
Weaknesses
- Hard to evaluate, because there is no single right answer.
- Results need human interpretation (what does cluster 3 mean?).
When to use itWhen you want to explore data, compress it, or find unusual points, and you have no labels.
Compare with
Check yourself
- Q01
Which task is unsupervised?
Show answer
B: Grouping news articles by topic without predefined topics