Introduced on day 14
Silhouette score
For each row, how much closer it is to its own cluster than to the nearest other cluster, scaled to the range -1 to 1 and averaged over rows.
In detail
Values near 1 mean tight, well-separated clusters; values near 0 mean overlapping clusters. It summarizes geometry for a given partition and distance, so it favours compact convex clusters and can be high for partitions driven by badly scaled features. It cannot tell real structure from the partition of a smooth cloud without a structureless reference and a stability check.