Unsupervised Learning — Beginner-Friendly Notes

1. What Unsupervised Means

Learn from unlabelled data $\{x_i\}$ with no teacher, no correct $y$. Goal: find hidden structure — groups, patterns, compression.

If data has no answers, and you still find meaning, it's unsupervised.

2. Two Main Families

Family	Goal	Example
Clustering	group similar points	customer segments, gene groups, image segments
Dimensionality reduction / Pattern discovery	compress, find hidden features	PCA, auto-encoders, association rules ("beer → diapers")

Clustering is the syllabus focus (k-means next topic).

3. How It Differs from Supervised

	Supervised	Unsupervised
Data	$(x,y)$ labelled	$x$ only
Task	predict $y$	discover structure
Evaluation	accuracy on test labels	subjective / indirect (coherence, downstream usefulness)
Example	classify spam (knows spam/ham)	group emails by topic (discovers topics)

4. Why Hard

No ground truth → hard to know if result is "correct". You evaluate by usefulness: tighter clusters (low within-cluster variance), human inspection, or better downstream task performance.

5. Typical Workflow

Unlabelled data → Similarity/Distance measure → Algorithm (k-means, hierarchical) → Clusters / embeddings → Inspect / validate

6. Applications

• Market segmentation, anomaly detection (fraud), gene clustering, document topic discovery, image segmentation.

7. One-Line Summary

Unsupervised learning extracts hidden structure — chiefly clusters or compressed representations — from unlabelled data, with evaluation by internal coherence and utility rather than label accuracy.

8. Quick Self-Check

1. Is grouping students by study habit without grades unsupervised?
2. Why is evaluation harder than supervised?