Don't Panic: Unsupervised Learning

Unsupervised learning is what happens when you hand an algorithm a table of observations and no answer key. No target label waits at the back of the book, wearing a tiny hat and declaring the winner. The method offers a view of structure: perhaps groups, unusual records, or fewer coordinates that retain useful variation.

The important bit is that the view depends on the map drawn first. An observation is one row or entity. A feature is a measurement that describes it. Decide those badly, and the algorithm will faithfully find patterns in the wrong geometry. It is not malicious. It is being extremely literal, which is often more troublesome.

Clustering is the familiar branch. K-means gathers points around chosen centers and asks for a cluster count. Density-based methods such as DBSCAN look for packed regions and can leave records as noise. Hierarchical clustering makes a tree of possible groupings. None receives an engraved invitation from reality confirming it found the one true partition. That invitation does not exist.

Dimensionality reduction changes the view rather than the rows. PCA finds linear directions of variation in centered data. It can reduce redundancy and make broad structure easier to inspect. A t-SNE plot can make local neighborhoods visible in two dimensions, which is useful, but it is not an authoritative atlas of every distance in the original data. Screen real estate has never been a reliable scientific instrument.

Without labels, evaluation needs more than one nod from a chart. The silhouette coefficient compares within-cluster closeness with closeness to another cluster. It helps compare candidates under a selected representation. It cannot establish that a cluster is stable, fair, causal, or useful. Repeat runs, samples, and domain review carry their share of the luggage.

Start with the Intro when the distinction between structure and prediction needs a foundation. Use Slides for method families and the decision path. Keep the Cheatsheet nearby when comparing assumptions and limits. The practice reference turns the whole business into a documented comparison, which is much harder to misplace than a colorful plot.