Classical ML, validation, metrics, interpretation, debugging · insight-0189
Test whether clusters survive resampling before naming them
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, Data analyst
- Industries
- retail, telecom
Summary: situation, action, reason, boundary
Situation
A k-means run on customer or route features produces k tidy segments that the business wants to name and target.
Action
Rerun with several seeds and bootstrap samples, measure agreement (adjusted Rand index between runs), and keep only segments that persist.
Reason
An algorithm's output is a partition by construction; only stability under perturbation suggests the partition reflects structure rather than the algorithm.
Boundary
k-means, Gaussian mixtures, hierarchical clustering and other segmentation used for business decisions.
Symptom or risk
k-means always returns k groups, even on structureless data, and different seeds or samples can rearrange them.
Exceptions
Clustering used only as compression (vector quantization) needs useful downstream performance, not stable meaning.
Trade-offs
Stability checks cost runs and may show fewer real segments than the business hoped for.
Pitfalls
Choosing k from the elbow of inertia alone, since inertia always decreases with k and is not normalized.
Lessons and concepts
Claims and their sources
k-means needs the number of clusters in advance, can converge to local minima depending on initialization, and its inertia is not a normalized metric.
Clustering results depend on choices such as scaling, dissimilarity and k, and clusters should be checked for robustness to perturbations of the data (ISL, clustering practical issues).