Stop 05 · Machine Learning FundamentalsTopics · 7/12

K-Means Clustering

In short

You have thousands of customers and no labels. K-Means groups similar ones together, so you can treat bargain hunters, weekend shoppers and big spenders differently.

An unsupervised algorithm that partitions data into K clusters by minimizing distance to cluster centroids — the default first tool for finding groups in unlabeled data.

The mascot holds a chai kettle on a lawn while three flags stand among three groups of tiny students.

Think of it like this

Placing chai stalls on campus

You may open three chai stalls on campus. Put them anywhere, then each student walks to the nearest one. Now shift each stall to the middle of its own crowd. Students re-pick their nearest stall. Repeat until nobody switches. Those three crowds are your clusters.

Watch it happen

One run of K-Means

Choose k

Drop centres

Assign

Move centres

Converged

1/6You decide how many clusters you want, say k = 3.

The big ideas

What you need to know

You must pick k

K-Means won't decide the number of groups. Try several values and use the elbow plot or silhouette score.

Scale your features

Distance drives everything. Income in rupees would drown out age in years unless you standardise first.

Likes round blobs

K-Means assumes compact, roughly round clusters of similar size. Long or curved shapes confuse it.

Start matters

Different starting centres can give different results. n_init runs it several times and keeps the best.

See it in code

Walk through the code

python

1from sklearn.cluster import KMeans2from sklearn.datasets import make_blobs3from sklearn.preprocessing import StandardScaler4 5X, _ = make_blobs(n_samples=300, centers=3, random_state=42)6X = StandardScaler().fit_transform(X)7 8for k in range(2, 7):9    km = KMeans(n_clusters=k, n_init=10, random_state=0).fit(X)10    print(k, round(km.inertia_, 1))   # look for the elbow11 12km = KMeans(n_clusters=3, n_init=10, random_state=0).fit(X)13print(km.cluster_centers_)

1/3Make 300 points in 3 hidden groups and scale them.

inertia_ is the total squared distance from points to their centre. It drops fast until the right k, then flattens.

Save yourself the pain

Common mistakes

Oops

Running K-Means on raw columns with wildly different units.

Do this instead

Standardise first, ideally with StandardScaler inside a Pipeline.

Oops

Treating cluster numbers 0, 1, 2 as meaningful labels.

Do this instead

Cluster IDs are arbitrary. Profile each cluster's averages and give it a human name yourself.

Your mission

Segment mall customers

~30 min

Quick check

After assigning points to centres, what does K-Means do next?

Go deeper

1Implement K-Means on a 2D dataset and visualize the clusters forming over iterations.
2Learn the elbow method for choosing K.
3Understand K-Means' limitations: it assumes spherical, similarly-sized clusters.