Statistics Glossary → Silhouette coefficient
Silhouette coefficientconcept
An index between -1 and +1 that measures how well each observation fits its own cluster compared with the nearest neighbouring cluster, summarizing clustering quality.
For an observation, a is the average distance to the other observations in its cluster and b is the average distance to the observations in the nearest other cluster. The silhouette value is s = (b - a) / max(a, b) (Rousseeuw). A value near 1 means the observation sits well in its cluster, a value around 0 means it lies on the border between two clusters, and a negative value means it is closer to another cluster, so it has probably been assigned to the wrong one.
The average silhouette over all observations is used to compare different numbers of clusters; the number giving the highest average stands out as a candidate. As a rough guide, above 0.7 suggests strong structure, 0.5-0.7 reasonable and 0.25-0.5 weak structure. The silhouette needs at least two clusters and tends to favour compact, similarly dense clusters; it can mislead for elongated, winding or unevenly dense clusters. It also depends on the chosen distance measure and on how the variables are scaled.
An everyday example
In a schoolyard, pupils have formed groups of friends. A pupil who stands very close to her own group and far from the others has a silhouette near 1. One standing exactly between two groups has a value around 0, and one standing far from his own group next to another group has a negative value.
In MerQur
Among the criteria for judging the number and quality of clusters in K-Means and Hierarchical Clustering; in DBSCAN Spatial Clustering, the 'Diagnostics' option gives the silhouette of the non-noise points.
Analyses where this concept appears
K-Means ClusteringHierarchical ClusteringDBSCAN Spatial Clustering
Related concepts
Choosing the number of clustersElbow method and k-distance plot