Skip to contents

A silhouette plot (Rousseeuw 1987) draws one bar per observation, grouped by cluster and sorted by width within each cluster, with the average width marked. plotSilhouette() draws it for any "Silhouette" object and for the silhouette results of cluster (Maechler et al. 2023), factoextra (Kassambara and Mundt 2020) and drclust (Prunila and Vichi 2025). It returns a ggplot2 object, so the plot can be changed and saved like any other.

library(Silhouette)
X <- as.matrix(iris[, 1:4])
set.seed(1)
km <- kmeans(X, centers = 3, nstart = 20)
sil <- Silhouette(proxy::dist(X, km$centers))

The default plot

plot() on a "Silhouette" object calls plotSilhouette(), so the two lines below draw the same plot.

plot(sil)

Silhouette plot of 150 iris flowers in three k-means clusters, with bars colored by cluster, a legend giving each cluster's size and average width, and a dashed red line at the overall average of about 0.67.

# plotSilhouette(sil)

The legend gives each cluster’s size and average width, the subtitle gives the overall average, and the dashed line marks it. The axis title names the width (medoid here). When the object was computed with average = "fuzzy" or average = "median", the subtitle says so and the line marks that average; the bars always show the individual widths.

Options

summary.legend = FALSE reduces the legend to the cluster numbers, grayscale = TRUE switches to gray bars and a black line for print, and linetype changes the line.

plotSilhouette(sil, summary.legend = FALSE, grayscale = TRUE, linetype = "solid")

The same silhouette plot in shades of gray with a solid black average line and a legend that lists only the cluster numbers.

Any further argument goes to ggpubr::ggpar(), which sets titles, axis labels, the legend position, the color palette, font sizes and the theme:

plotSilhouette(sil,
  title = "k-means, K = 3", ylab = "Medoid width",
  legend = "bottom", palette = c("#2a78d6", "#eb6834", "#1f9d74"))

The silhouette plot with a custom title, the legend below the plot, and bars in blue, orange and green.

The palette colors the clusters and the summary legend is kept. With grayscale = TRUE the palette is ignored.

Because the result is a ggplot, layers and themes can also be added with +, and ggplot2::ggsave() saves it to a file:

p <- plotSilhouette(sil) + ggplot2::theme(legend.position = "top")
ggplot2::ggsave("silhouette.pdf", p, width = 7, height = 4)

Finding the poorly placed observations

With label = TRUE the bars are labeled with the row index of each observation. This is readable for up to about 50 observations. For the USArrests data, with four k-means clusters of the standardized crime rates:

Xu <- scale(USArrests)
set.seed(1)
km_us <- kmeans(Xu, centers = 4, nstart = 20)
sil_us <- Silhouette(proxy::dist(Xu, km_us$centers))
plotSilhouette(sil_us, label = TRUE)

Silhouette plot of 50 US states in four clusters with row indices under the bars. A few states have widths below 0.2.

The object keeps the rows in the order of the data, so the widths can be matched back to the observations directly:

rownames(USArrests)[sil_us$sil_width < 0.2]
#> [1] "Alaska"    "Kentucky"  "Minnesota" "Missouri"  "Nebraska"

Negative widths

When the labels come from a membership matrix passed as prob_matrix, or from the cluster argument, an observation whose assigned cluster is not its nearest cluster center gets a negative width, and its bar points downward. For a Gaussian mixture of the iris data, scored against the component means:

suppressPackageStartupMessages(library(mclust))
gm <- Mclust(X, G = 3, verbose = FALSE)
sil_gm <- Silhouette(proxy::dist(X, t(gm$parameters$mean)), prob_matrix = gm$z)
plotSilhouette(sil_gm)

Silhouette plot of the iris data in three Gaussian mixture components, with widths computed from distances to the component means. A handful of observations in two clusters have negative widths.

These are the observations on which the mixture and the geometry disagree; see vignette("interpreting-widths", package = "Silhouette").

Results from other packages

plotSilhouette() also draws the silhouette results of partitioning around medoids and its relatives in cluster (pam(), clara(), fanny() and silhouette()), of eclust() and hcut() in factoextra, and of silhouette() in drclust. These packages compute the classical widths from all pairwise distances; plotSilhouette() only draws them. Call plotSilhouette() explicitly for these objects, because plot() dispatches to the methods of the package that created them.

pm <- cluster::pam(X, k = 3)
fy <- cluster::fanny(X, k = 3)
ec <- factoextra::eclust(X, "kmeans", k = 3, graph = FALSE)
hc <- factoextra::hcut(X, k = 3)
small <- function(x, title) plotSilhouette(x, title = title, font.title = 10,
                                           font.subtitle = 8, font.legend = 7)
ggpubr::ggarrange(
  small(pm, "cluster::pam()"),
  small(fy, "cluster::fanny()"),
  small(ec, "factoextra::eclust()"),
  small(hc, "factoextra::hcut()"),
  ncol = 2, nrow = 2)

Four silhouette plots of the iris data in a two-by-two grid, for pam, fanny, factoextra eclust with k-means, and factoextra hcut. All four show one well-separated cluster and two overlapping ones.

For fanny(), the widths are those of the crisp partition obtained by assigning each observation to its cluster of largest membership. The same holds for objects returned by cluster::silhouette(), for example for a hierarchical clustering:

hc_avg <- cutree(hclust(dist(X), method = "average"), k = 3)
plotSilhouette(cluster::silhouette(hc_avg, dist(X)))

Silhouette plot of the iris data in three average-linkage clusters, with classical widths from cluster::silhouette.

In drclust, silhouette() takes the data and a fitted model, such as the reduced k-means of dpcakm() with \(K\) clusters in \(Q\) components, and returns the classical widths in component cl.silhouette:

set.seed(1)
dp <- drclust::dpcakm(X, K = 3, Q = 2)
d <- drclust::silhouette(X, dp)
#>   cluster size ave.sil.width
#> 1       1   52          0.39
#> 2       2   50          0.79
#> 3       3   48          0.32
plotSilhouette(d$cl.silhouette)

Silhouette plot of the iris data in three clusters from drclust dpcakm, with classical widths.

drclust::silhouette() prints a summary table when it is called; the drclust package also masks cluster::silhouette() when it is attached, so the calls above use ::.

Widths computed elsewhere

getSilhouette() builds a "Silhouette" object from widths computed by any other means, so that summary() and plot() work on them. It needs the assigned cluster, the neighboring cluster and the width of every observation; weight is needed only for fuzzy averaging. is.Silhouette() checks the result, and with strict = TRUE also its columns and attributes.

For example, approxSilhouette() in the Bioconductor package bluster (Lun 2023) approximates the classical widths of a crisp partition in linear time and returns them as a DataFrame:

a <- bluster::approxSilhouette(X, pm$clustering)
sil_bl <- getSilhouette(cluster = as.integer(a$cluster),
                        neighbor = as.integer(a$other),
                        sil_width = a$width)
is.Silhouette(sil_bl, strict = TRUE)
#> [1] TRUE
summary(sil_bl, print.summary = FALSE)$avg.width
#> [1] 0.5403281
plotSilhouette(sil_bl)

Silhouette plot of approximate classical widths from bluster for the pam partition of the iris data.

The same pattern works for widths computed in other software and read into R.

Session information

sessionInfo()
#> R version 4.6.2 (2026-10-29)
#> Platform: x86_64-pc-linux-gnu
#> Running under: Ubuntu 26.04.1 LTS
#> 
#> Matrix products: default
#> BLAS:   /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3 
#> LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.32.so;  LAPACK version 3.12.0
#> 
#> locale:
#>  [1] LC_CTYPE=C.UTF-8       LC_NUMERIC=C           LC_TIME=C.UTF-8       
#>  [4] LC_COLLATE=C.UTF-8     LC_MONETARY=C.UTF-8    LC_MESSAGES=C.UTF-8   
#>  [7] LC_PAPER=C.UTF-8       LC_NAME=C              LC_ADDRESS=C          
#> [10] LC_TELEPHONE=C         LC_MEASUREMENT=C.UTF-8 LC_IDENTIFICATION=C   
#> 
#> time zone: UTC
#> tzcode source: system (glibc)
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> other attached packages:
#> [1] mclust_6.1.3     Silhouette_0.9.7
#> 
#> loaded via a namespace (and not attached):
#>  [1] gtable_0.3.6           xfun_0.62              bslib_0.12.0          
#>  [4] ggplot2_4.0.4          htmlwidgets_1.6.4      ggrepel_0.9.8         
#>  [7] rstatix_1.1.0          lattice_0.23-1         vctrs_0.7.3           
#> [10] tools_4.6.2            generics_0.1.4         stats4_4.6.2          
#> [13] flexmix_2.3-21         parallel_4.6.2         tibble_3.3.1          
#> [16] proxy_0.4-29           DEoptimR_1.2-2         cluster_2.1.8.3       
#> [19] BiocNeighbors_2.6.0    pkgconfig_2.0.3        Matrix_1.7-6          
#> [22] pheatmap_1.0.13        RColorBrewer_1.1-3     S4Vectors_0.50.3      
#> [25] S7_0.2.2               desc_1.4.3             lifecycle_1.0.5       
#> [28] factoextra_2.2.0       compiler_4.6.2         farver_2.1.2          
#> [31] textshaping_1.0.5      bluster_1.22.0         codetools_0.2-20      
#> [34] carData_3.0-6          htmltools_0.5.9        class_7.3-24          
#> [37] sass_0.4.10            yaml_2.3.12            Formula_1.2-6         
#> [40] pillar_1.11.1          pkgdown_2.2.1          car_3.1-5             
#> [43] ggpubr_1.0.0           jquerylib_0.1.4        tidyr_1.3.2           
#> [46] prabclus_2.3-5         MASS_7.3-66            BiocParallel_1.46.0   
#> [49] diptest_0.77-2         cachem_1.1.0           fpc_2.2-15            
#> [52] abind_1.4-8            robustbase_0.99-7      tidyselect_1.2.1      
#> [55] drclust_0.1.1          digest_0.6.39          dplyr_1.2.1           
#> [58] kernlab_0.9-33         purrr_1.2.2            labeling_0.4.3        
#> [61] RcppArmadillo_15.6.2-1 cowplot_1.2.0          fastmap_1.2.0         
#> [64] grid_4.6.2             cli_3.6.6              magrittr_2.0.5        
#> [67] broom_1.0.13           withr_3.0.3            scales_1.4.0          
#> [70] backports_1.5.1        rmarkdown_2.32         igraph_2.3.4          
#> [73] otel_0.3.0             nnet_7.3-21            ggsignif_0.6.4        
#> [76] ragg_1.5.2             modeltools_0.2-25      evaluate_1.0.5        
#> [79] knitr_1.52             rlang_1.3.0            Rcpp_1.1.2            
#> [82] glue_1.8.1             BiocGenerics_0.58.1    jsonlite_2.0.0        
#> [85] R6_2.6.1               systemfonts_1.3.2      fs_2.1.0

References

Kassambara, Alboukadel, and Fabian Mundt. 2020. factoextra: Extract and Visualize the Results of Multivariate Data Analyses. https://CRAN.R-project.org/package=factoextra.
Lun, Aaron. 2023. bluster: Clustering Algorithms for Bioconductor. https://doi.org/10.18129/B9.bioc.bluster.
Maechler, Martin, Peter Rousseeuw, Anja Struyf, and Mia Hubert. 2023. cluster: “Finding Groups in Data”: Cluster Analysis Extended Rousseeuw Et Al. https://CRAN.R-project.org/package=cluster.
Prunila, Ionel, and Maurizio Vichi. 2025. “drclust: An R Package for Simultaneous Clustering and Dimensionality Reduction.” The R Journal 17 (4): 104–33. https://doi.org/10.32614/RJ-2025-046.
Rousseeuw, Peter J. 1987. “Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis.” Journal of Computational and Applied Mathematics 20: 53–65. https://doi.org/10.1016/0377-0427(87)90125-7.