Metrics daviesbouldin

DaviesBouldin

For each cluster, the worst ratio of "how spread these two are" to "how far apart they sit", averaged over the clusters. Lower is better — the opposite direction to CalinskiHarabasz and Silhouette, and 0 is the floor.

That inversion is the one thing worth checking before reading a table of these: a clustering that improves moves this number down and the other two up, and a reader who takes the three as interchangeable will read one of them backwards.

Like CalinskiHarabasz, it scores a clustering against nothing but the samples, takes a feature matrix rather than two label vectors, and has no precomputed-distance form — both read cluster centroids, which a distance matrix does not carry.

Two clusters sharing a centroid contribute nothing

The ratio would divide by zero. The reference replaces a zero centroid distance with infinity before dividing, so the pair scores 0 and drops out of its cluster's worst case. Measured, four identical points split into two clusters score 0 — as do two well-separated points each duplicated into a cluster of its own, which is a perfect clustering. The floor is reached from both directions, and the number cannot tell them apart.

Members

Member	What it does
DaviesBouldin.Score	The mean worst-case similarity between a cluster and any other.