Clustering with K-Means 🔒 Premium

Finding groups nobody labeled

No churn column, no spam flags — just customers described by spending, visits, tenure. Clustering discovers natural groupings anyway: bargain hunters, loyal regulars, about-to-leave. It's exploration, not prediction — the output is insight ("oh, there are four kinds of users"), which then drives decisions, features, or targeted models per group.

How K-means thinks

Pick K center points. Repeat: assign every point to its nearest center, then move each center to its group's middle. Stop when centers stop moving. The result: K blobs minimizing within-group distance.

from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans

Xs = StandardScaler().fit_transform(X)  # scale FIRST — see below
labels = KMeans(n_clusters=4, n_init=10, random_state=7).fit_predict(Xs)

The two things beginners get wrong

• Unscaled features. Income (tens of thousands) vs visits (single digits): raw distance is 99% income. Standardize first, always, for any distance-based method.
• Treating K as truth. K-means always returns K groups, even from pure noise. Try several K values, elbow-plot the inertia (the "bend" suggests a natural count), and — critically — validate clusters mean something in the business, not just the math. A segment nobody can act on is decoration.

Remember this

• No labels → cluster for structure; scale first; choose K by elbow + business sense.
• Clusters are hypotheses to validate, not answers to ship blindly.