Unlock Hidden Patterns: A Hands‑On Guide to Unsupervised Learning Techniques

What Is Unsupervised Learning?

Unsupervised learning extracts hidden patterns from data that has no labels, typically using clustering, dimensionality reduction, or association algorithms to discover underlying regularities.

Key traits: no teacher guidance, data lacks explicit labels. Common misconception: it does not mean humans are completely absent; algorithmic objectives still need to be designed.

How to Implement

The three main families of unsupervised methods are:

Clustering – groups similar data points.

Dimensionality reduction – compresses feature space.

Association – discovers co‑occurrence rules.

Clustering (K‑means) – Grouping Similar Items

Problem addressed: “Which things are intrinsically similar?”

Example: automatically partitioning 200 dishes in a buffet.

Original state: 200 dishes scattered.

Clustering process:

✓ Algorithm extracts dish features (cooking method, ingredients, taste).

✓ Automatically creates groups: seafood/sashimi zone, spicy Sichuan zone, Western bakery zone.

Value: customers can locate their desired area within 5 seconds.

Dimensionality Reduction (PCA) – Simplifying Complex Information

Problem addressed: “How can we simplify high‑dimensional data?”

Example: a housing‑decision model with 20 original dimensions (school district, commute time, green space, property fees, …).

Reduction process:

✓ Algorithm extracts core features, producing an education‑resource index and a livability score.

✓ Generates a 2‑D map for visual inspection.

Value: users can pinpoint a target house in about half an hour.

Association – Discovering Co‑Occurrence Rules

Problem addressed: “Which events tend to happen together?”

Example: product placement strategy in a convenience store.

Original data: 100,000 transaction receipts.

Association rule mining:

{chips, soda} → {tissue} [support = 22 %, confidence = 81 %].

Interpretation: 81 % of customers buying snacks and drinks also pick up a tissue.

Value: placing a tissue rack next to the checkout increased tissue sales by 35 %.

Hands‑On Experiments

Clustering demo: use K-means GUI to visualize the grouping process (online view).

Dimensionality‑reduction demo: view word‑vector compression with TensorFlow Embedding Projector (online view).

Association demo: implement a Python script to analyse timeout‑cart data (online view).

Real‑World Cases

Amazon employs DeepCluster to split products into 270,000 clusters, achieving 19× more granularity than manual categorisation.

NASA uses t-SNE to analyse galaxy images, cutting processing time from three weeks to four hours.

7‑Eleven discovers the rule 关东煮 + 清酒 → increased winter sales by $6,800 per store.

References

“Unsupervised Learning” – 一文学会无监督学习.