Clustering Notes
Companion page for the k-means and k-means++ notebooks — patterned segmentation for people who already speak GLM.
Clustering is unsupervised storytelling: you propose a grammar for “similar,” then listen for whether the data sings in key.
Where to read it
k-means clustering
Classical centroid updates, inertia, and the usual caveats about scale and spherical assumptions.
k-means++
Smarter seeding to reduce unlucky local minima — a small algorithmic kindness with big practical payoff.
How to read clustering as a predictive person
- Scaling first. Distance-based methods inherit your measurement units. Standardize on purpose.
- k is a modeling choice. Elbow plots help; domain cost of a segment helps more.
- Stability over spectacle. Bootstrap the assignment. If labels reshuffle wildly, the story is soft.
- Downstream use. Clusters become strata for GLMs, marketing treatments, or reserving cohorts — judge them by that job.
The worked comparison of clustering methods is in the Clustering repository.
The study manual chapter is the on-site walkthrough.