Unsupervised Learning: Clustering, Dimensionality Reduction

Unsupervised learning works with data that has no labels — no correct answers to learn from. Instead, the algorithm discovers structure, patterns, and groupings on its own. It answers questions like 'what natural groups exist in this data?' and 'which features matter most?'

This lesson covers the two primary unsupervised learning tasks: clustering (grouping similar data points) and dimensionality reduction (simplifying complex data while preserving its essence).

Clustering algorithms for unsupervised learning

Clustering groups data points that are similar to each other:

K-Means — the most popular. Partition data into K clusters. Fast and scalable. You choose K (number of clusters) upfront.
DBSCAN — density-based clustering. Finds clusters of arbitrary shape and identifies outliers. No need to specify K.
Hierarchical clustering — builds a tree of clusters. Useful for small datasets where you want to see the cluster hierarchy (dendrogram).

Use K-Means for most tasks. Choose K using the elbow method (plot inertia vs K, pick the 'elbow' point) or silhouette score. DBSCAN when clusters have irregular shapes or when outlier detection matters.

Dimensionality reduction techniques

Real-world data often has dozens or hundreds of features. Dimensionality reduction simplifies it:

PCA (Principal Component Analysis) — projects data onto the directions of maximum variance. Reduces 100 features to 2–10 principal components while keeping most information. The standard technique.
t-SNE — non-linear reduction optimised for visualization. Maps high-dimensional data to 2D or 3D for plotting. Great for exploring clusters visually.
UMAP — faster than t-SNE with better preservation of global structure. Increasingly preferred for visualization tasks.

PCA is used both for visualization and as a preprocessing step before training ML models. Fewer dimensions means faster training and less overfitting.

Practical applications of unsupervised learning

Unsupervised learning powers many real-world systems:

Customer segmentation — group customers by behaviour for targeted marketing
Anomaly detection — find fraudulent transactions, network intrusions, or manufacturing defects
Topic modelling — discover themes in large document collections
Feature engineering — create meaningful groupings as inputs for supervised models
Data exploration — understand the natural structure of a new dataset before building models

Unsupervised learning is often the first step in a larger pipeline. Cluster your data, understand the segments, then build supervised models for each segment — this often outperforms a single global model.

When evaluating unsupervised learning, start with the fundamentals: what problem it solves, who it serves, and how it compares to the alternatives buyers consider first.

Start with the fundamentals. Define who you are serving, the single outcome you want, and the budget you are comfortable committing before you compare any options.

A practical rule is to favour clarity over cleverness. The choice that is easiest to understand, spell, remember and act on tends to win over the long run.

Context changes the answer. What is optimal for a fast-moving startup is rarely optimal for an established brand, so weigh your own stage, market and ambitions.

Avoid the common mistakes: rushing the decision, copying competitors blindly, and ignoring the total cost of ownership beyond the very first year of use.

Think in terms of trade-offs rather than a single perfect answer. Every option trades memorability, cost, reach and flexibility in a slightly different mix.

Document your reasoning. Writing down why you chose one path makes it far easier to revisit the decision later and to explain it to partners and investors.

Test with real people. A quick check with a handful of your actual customers will surface problems that no amount of internal debate ever reveals.

Plan for growth. The option that fits today should still make sense when your audience, catalogue or team is several times larger than it is right now.

Finally, revisit the decision periodically. Markets, tools and best practices move quickly, and a choice that was ideal last year may deserve a fresh look today.