Stage 2 · Intermediate · M4
Finding groups nobody labeled
The model has no answers to copy. It has to invent the categories.
12 lessons · 112 minSteady
About this chapter
With no labels, a model still has one move: measuring how far apart two things are. That one move builds everything here. You will run k-means by hand until it stops, see the shapes it gets wrong, cut a tree of merges at a sensible height, and meet a method that is allowed to answer nowhere. Then the chapter stops grouping and starts squeezing, from a rotation that keeps the widest view to a grid of scores rebuilt from two thin tables. It ends by burying real groups under columns that mean nothing.
What you will be able to do
- 18 min
Near means similar
Say why distance is the one tool left without labels, and pick the scale that decides it.
- 28 min
Who is nearest
Find a point's nearest neighbors with no labels, and read how crowded its spot is.
- 310 min
k-means, step by step
Run assign and update by hand until the centers stop moving.
- 411 min
Where k-means breaks
Spot a bad landing, and predict k-means failing on crescents, rings, long stripes and lopsided groups.
- 58 min
How many groups are there?
Use elbow and silhouette to argue for a value of k, and admit the ambiguity.
- 69 min
Clusters inside clusters
Read a dendrogram and cut it at a sensible height.
- 79 min
Density, not centers
Tune DBSCAN's two parameters and read which points it calls noise.
- 811 min
PCA is a rotation
Turn the data to its widest directions and read how much of the spread each one keeps.
- 910 min
Maps of high-dimensional data
Read a t-SNE or UMAP map and list what distances on it do not mean.
- 109 min
Points that belong nowhere
Score points for strangeness, catch the odd combination one column misses, and set a line by how many alerts people can check.
- 1110 min
A grid with less in it
Rebuild a full grid of scores from a few numbers per row and per column, and say what those numbers are.
- 129 min
The curse of dimensionality
Explain why columns that mean nothing bury real groups, and what people do before they cluster.