Inertia—Lower is better
Silhouette—Higher is better
Davies–Bouldin—Lower is better
Calinski–Harabasz—Higher is better
Gap statistic—Higher is better
Cluster balance—Higher is better
Stability—Higher is better
Runtime—Milliseconds
Tiny clusters—Fewer is better
Isolated points—Potential outliers
Visual Evidence Dashboard
Cluster Count Comparison
| Pin | K | Inertia | Silhouette | DBI | CH | Gap | Balance | Stability | Runtime | Status |
|---|
Coach, History, and Exports
Generate a dataset, run one cluster count, then compare the full range.
Achievements: none yet.
Learning Notes
Elbow method
Look for a sharp reduction in improvement. The elbow marks diminishing returns. Ambiguous curves require other evidence.
Silhouette analysis
Values near one indicate separation. Values near zero suggest overlap. Negative values suggest questionable assignments.
Under-clustering
Too few clusters merge distinct patterns. Inertia stays high. Visual groups may contain several dense regions.
Over-clustering
Too many clusters split useful groups. Tiny clusters appear. Interpretability and stability often decline.
Stability
A useful solution survives different initialisations. Unstable labels suggest weak structure. Compare agreement across repeated runs.
Practical choice
The mathematically strongest K is not always best. Prefer understandable, useful, balanced, and operationally affordable segments.