Average silhouette evaluates candidate cluster counts by comparing each observation’s cohesion within its cluster with its separation from the nearest alternative cluster. Each silhouette coefficient ranges from −1 to +1: positive values indicate better-separated assignments, while low values suggest that K may be too low or too high. section 11.5.2
Unlike the elbow method, which depends on identifying a bend in a scree plot, the silhouette method provides a more precise choice of K. Plot the candidate cluster counts against their silhouette coefficients and select the count with the strongest result. section 11.5.2 section 20.1
Choosing clusters for NBA teams
For standardized team payroll and regular-season wins from 2000–2017, factoextra’s fviz_nbclust() compares cluster counts using the silhouette method. The silhouette result returned two clusters, while the within-cluster sum of squares method returned six; the analysis ultimately used four clusters as a compromise. section 11.5.2 section 20.1
options(scipen = 000) p3 <- fviz_nbclust(final_kmeans, kmeans, method = "wss") p4 <- fviz_nbclust(final_kmeans, kmeans, method = "silhouette")
Use the R listing to generate the within-cluster sum of squares and silhouette diagnostics with fviz_nbclust(). section 11.5.2
