Skip to contents

This article compares three soft clusterings of the same patients with the widths of Silhouette, and then uses the known diagnoses to check what the widths reveal. It follows the thyroid example of the package paper and shows the full code.

The UCI thyroid data, distributed with mclust (Scrucca et al. 2016), record five laboratory measurements on 215 patients whose thyroid function was diagnosed from their full medical record as normal, hypothyroid or hyperthyroid.

library(Silhouette)
suppressPackageStartupMessages(library(mclust))
data(thyroid, package = "mclust")
table(thyroid$Diagnosis)
#> 
#>   Hypo Normal  Hyper 
#>     30    150     35
Xt <- scale(thyroid[, -1])   # standardized laboratory measurements

Three soft clusterings

We fit three methods with \(K = 3\): fuzzy c-means (Bezdek 1981) from ppclust, a Gaussian mixture from mclust whose covariance structure is chosen by BIC, and probabilistic distance clustering adjusted for cluster size (PDQ, Iyigun and Ben-Israel 2008) from FPDclustering (Tortora and Palumbo 2025). The diagnoses are not used in fitting.

set.seed(1)
fc <- ppclust::fcm(Xt, centers = 3, nstart = 5)
gm <- Mclust(Xt, G = 3, verbose = FALSE)
pq <- FPDclustering::PDQ(Xt, k = 3)

For each fit we need two \(215 \times 3\) matrices: the Euclidean distances from each patient to each cluster center, and the membership or posterior probabilities. ppclust::fcm() stores squared distances, so we take square roots.

fits <- list(
  FCM = list(prox = sqrt(fc$d), prob = fc$u),
  GMM = list(prox = as.matrix(proxy::dist(Xt, t(gm$parameters$mean))), prob = gm$z),
  PDQ = list(prox = as.matrix(proxy::dist(Xt, pq$centers)), prob = pq$probability))
labels <- lapply(fits, function(f) max.col(f$prob))   # the partition each method produced

Average widths

For the distance-based widths we pass the memberships as prob_matrix, so that every patient keeps the cluster the method assigned (passing cluster = labels[[...]] would do the same). For the probability-based widths we pass the memberships alone. The density-based width warns about zero posteriors, which are discussed below, so its warnings are suppressed here.

avg <- function(s) summary(s, print.summary = FALSE)$avg.width
tab <- sapply(fits, function(f) c(
  medoid     = avg(Silhouette(f$prox, prob_matrix = f$prob)),
  pac        = avg(Silhouette(f$prox, prob_matrix = f$prob, method = "pac")),
  PACS       = avg(softSilhouette(f$prob)),
  NLPPS      = avg(softSilhouette(f$prob, prob_type = "nlpp", method = "medoid")),
  CeS        = avg(cerSilhouette(f$prob)),
  DBS        = avg(suppressWarnings(dbSilhouette(f$prob, average = "crisp")))))
ari <- sapply(labels, adjustedRandIndex, thyroid$Diagnosis)
knitr::kable(rbind(tab, ARI_with_diagnosis = ari), digits = 3)
FCM GMM PDQ
medoid 0.659 0.588 -0.026
pac 0.521 0.460 -0.018
PACS 0.764 0.975 0.429
NLPPS 0.879 0.985 0.703
CeS 0.839 0.987 0.688
DBS 0.466 0.085 0.291
ARI_with_diagnosis 0.659 0.892 0.015

The two families of widths rank the methods differently. The distance-based widths rank them FCM, GMM, PDQ, and the probability-based widths rank them GMM, FCM, PDQ. Agreement with the diagnosis, which would not be available in practice, orders them GMM, FCM, PDQ.

The difference is expected. Distances to cluster means reward the spherical, equal-spread clusters that fuzzy c-means looks for, while the mixture (here with covariance model VVI) allows clusters of different shapes; posterior probabilities reward decisive assignment, which the mixture achieves. In this example the probability-based widths pointed to the partition that matches the diagnosis, but neither family is a substitute for knowing what kind of cluster the analysis is after.

The distance-based and probability-based widths agree on PDQ: every one of them is lowest for it. With the default initialization, PDQ placed 214 of the 215 patients in one cluster; five random initializations gave similar partitions, all with an adjusted Rand index below 0.05.

The density-based width is the exception, and it is lowest for the mixture. Each density-based width is divided by the largest finite log ratio in the sample. For the mixture that log ratio is about 691, against 5.3 for fuzzy c-means, so most of the mixture’s widths are divided by a very large number. The density-based average for the mixture therefore reflects its few most confident patients more than the partition; see the section on zero posteriors below.

The same table, with fuzzy and median averages as well, comes from calSilhouette(), which scores the partition given by the memberships in every row:

cs <- suppressWarnings(calSilhouette(prox_matrix = fits$GMM$prox, prob_matrix = fits$GMM$prob))
cs
#>         Method      Crisp      Fuzzy     Median
#> 1       medoid 0.58765397 0.60352421 0.67490839
#> 2          pac 0.46029717 0.47335031 0.50932960
#> 3       pp_pac 0.97487592 0.99291881 0.99998172
#> 4    pp_medoid 0.98060299 0.99602933 0.99999086
#> 5     nlpp_pac 0.98147322 0.99713635 0.99999842
#> 6  nlpp_medoid 0.98500675 0.99840423 0.99999921
#> 7       pd_pac 0.97446673 0.98514014 0.99992183
#> 8    pd_medoid 0.98375861 0.99087405 0.99996092
#> 9          cer 0.98743592 0.99645732 0.99999086
#> 10          db 0.08523792 0.08822245 0.01678418
all.equal(cs$Crisp[cs$Method == "medoid"], unname(tab["medoid", "GMM"]))
#> [1] TRUE

Scoring the distances without the memberships would assign every patient to the nearest center instead, a different partition for these fits:

sapply(fits, function(f) sum(max.col(f$prob) != max.col(-f$prox)))
#> FCM GMM PDQ 
#>   0  14 115

Which patients are poorly clustered?

The individual widths are more useful than the averages. To check them, we map each cluster to its majority diagnosis and mark the patients whose diagnosis differs.

wrong <- lapply(labels, function(l) {
  majority <- apply(table(l, thyroid$Diagnosis), 1, which.max)
  levels(thyroid$Diagnosis)[majority[as.character(l)]] != as.character(thyroid$Diagnosis)
})
sapply(wrong, sum)
#> FCM GMM PDQ 
#>  22   7  64
w_med <- lapply(fits, function(f) Silhouette(f$prox, prob_matrix = f$prob)$sil_width)
w_pp <- lapply(fits, function(f) softSilhouette(f$prob)$sil_width)
st <- do.call(rbind, lapply(names(fits), function(nm) rbind(
  data.frame(method = nm, panel = "distance-based (medoid)", width = w_med[[nm]],
             wrong = wrong[[nm]]),
  data.frame(method = nm, panel = "probability-based (pp, PAC)", width = w_pp[[nm]],
             wrong = wrong[[nm]]))))
st$status <- factor(ifelse(st$wrong, "disagrees with diagnosis", "agrees with diagnosis"),
                    c("agrees with diagnosis", "disagrees with diagnosis"))
st <- st[order(st$wrong), ]   # draw disagreeing patients on top

library(ggplot2)
set.seed(3)
ggplot(st, aes(method, width, colour = status, shape = status)) +
  geom_jitter(width = 0.22, height = 0, size = 1.4, alpha = 0.85) +
  facet_wrap(~ panel) +
  scale_colour_manual(values = c("#2a78d6", "#eb6834"), name = NULL) +
  scale_shape_manual(values = c(16, 17), name = NULL) +
  labs(x = NULL, y = "Silhouette width") +
  theme_minimal(base_size = 10) +
  theme(legend.position = "bottom")

Two strip charts of per-patient widths for FCM, GMM and PDQ. In the distance-based panel, patients whose cluster disagrees with the diagnosis sit mostly at low or negative widths for FCM and GMM. In the probability-based panel, GMM widths are almost all near 1, including most disagreeing patients. PDQ widths are low throughout.

Because the labels come from the memberships, a patient whose assigned cluster is not the nearest center has a negative medoid width. To compare the widths with the diagnoses, we take the tenth of patients with the smallest widths and compare the share of misassigned patients among them with the share among the rest:

low_share <- function(w, bad) {
  low <- w <= quantile(w, 0.1)
  c(lowest_tenth = mean(bad[low]), others = mean(bad[!low]))
}
round(rbind(
  medoid_FCM = low_share(w_med$FCM, wrong$FCM),
  medoid_GMM = low_share(w_med$GMM, wrong$GMM),
  pp_GMM = low_share(w_pp$GMM, wrong$GMM)), 2)
#>            lowest_tenth others
#> medoid_FCM         0.36   0.07
#> medoid_GMM         0.14   0.02
#> pp_GMM             0.14   0.02

By rank, both families point to the same patients: misassigned patients are concentrated among the smallest widths. In absolute terms they do not. For the mixture, the misassigned patients’ probability-based widths are still high,

c(misassigned = median(w_pp$GMM[wrong$GMM]), others = median(w_pp$GMM[!wrong$GMM]))
#> misassigned      others 
#>   0.9932631   0.9999855

so the mixture is confidently wrong about them, whereas their medoid widths place them among the poorly separated patients:

c(misassigned = median(w_med$GMM[wrong$GMM]), others = median(w_med$GMM[!wrong$GMM]))
#> misassigned      others 
#>   0.4513205   0.6865956

In practice, the patients with the smallest widths are the ones to review first. The object keeps the rows in the order of the data, so they are easy to list:

sil_gmm <- Silhouette(fits$GMM$prox, prob_matrix = fits$GMM$prob)
worst <- order(sil_gmm$sil_width)[1:8]
data.frame(patient = worst, cluster = sil_gmm$cluster[worst],
           width = round(sil_gmm$sil_width[worst], 3),
           diagnosis = thyroid$Diagnosis[worst])
#>   patient cluster  width diagnosis
#> 1      14       3 -0.839    Normal
#> 2     211       3 -0.390      Hypo
#> 3      50       3 -0.369    Normal
#> 4     178       1 -0.345     Hyper
#> 5     214       3 -0.335      Hypo
#> 6     185       1 -0.293     Hyper
#> 7     152       1 -0.218     Hyper
#> 8     153       1 -0.192     Hyper

The density-based width and zero posteriors

The density-based width of Menardi (2011) scales each log ratio \(\log(\gamma_{ic_i}/\gamma_{i\nu_i})\) by the largest one in the sample. For some patients the mixture’s posterior probability of the neighboring component underflows to exactly zero:

sum(apply(gm$z, 1, function(z) sort(z, decreasing = TRUE)[2]) == 0)
#> [1] 7

Their log ratios are infinite. dbSilhouette() gives these patients width 1, scales the other widths by the largest finite log ratio, and says so in a warning. (Up to version 0.9.6, the infinite maximum made every other width zero.)

db_gmm <- dbSilhouette(gm$z, average = "crisp")
#> Warning: 7 observation(s) have a zero membership in the neighboring (or
#> assigned) cluster, so their log ratio is infinite; their widths are set to 1
#> (or -1) and the other widths are scaled by the largest finite log ratio.
avg(db_gmm)
#> [1] 0.08523792

The average is still small, because the largest finite log ratios of the mixture are huge and set the scale for everyone else. The individual density-based widths still rank the patients by how decisively they are assigned, but the average of a mixture with near-zero posteriors should not be compared with that of a method without them.

Fuzzy averaging

With average = "fuzzy", each width is weighted by \((\gamma_{i(1)} - \gamma_{i(2)})^a\), the gap between the two largest memberships (Campello and Hruschka 2006). For the fuzzy c-means memberships the average depends clearly on the exponent:

sapply(c(a0.5 = 0.5, a1 = 1, a2 = 2, a4 = 4), function(a)
  avg(softSilhouette(fc$u, average = "fuzzy", a = a)))
#>      a0.5        a1        a2        a4 
#> 0.8077759 0.8344067 0.8668654 0.8996649

Larger exponents shift weight toward decisively assigned patients, so fuzzy averages are comparable only at the same a.

Lessons

  • Compute each width for every fit and compare fits within a width, never widths with each other.
  • Distance-based and probability-based widths answer different questions and can rank methods differently. Report both when the methods assume different cluster shapes.
  • A method that fails, like PDQ here, is flagged by the distance-based and the probability-based widths alike.
  • Use the individual widths to find the observations to review, and pass the memberships as prob_matrix (or the labels as cluster) so that the widths refer to the partition the method produced.
  • Read the density-based average with care when posteriors are near zero: the most extreme patients set its scale.

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] ggplot2_4.0.4    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] klaR_1.7-4          htmlwidgets_1.6.4   rstatix_1.1.0      
#>  [7] GGally_2.5.0        vctrs_0.7.3         tools_4.6.2        
#> [10] generics_0.1.4      tibble_3.3.1        proxy_0.4-29       
#> [13] highr_0.12          cluster_2.1.8.3     pkgconfig_2.0.3    
#> [16] prettyGraphs_2.2.0  kpeaks_1.1.0        RColorBrewer_1.1-3 
#> [19] S7_0.2.2            desc_1.4.3          lhs_1.3.0          
#> [22] lifecycle_1.0.5     rootSolve_1.8.2.4   farver_2.1.2       
#> [25] compiler_4.6.2      inaparc_1.2.1       textshaping_1.0.5  
#> [28] combinat_0.0-9      ppclust_1.1.0.1     carData_3.0-6      
#> [31] httpuv_1.6.17       htmltools_0.5.9     sass_0.4.10        
#> [34] yaml_2.3.12         Formula_1.2-6       car_3.1-5          
#> [37] ggpubr_1.0.0        tidyr_1.3.2         later_1.4.8        
#> [40] pkgdown_2.2.1       pillar_1.11.1       jquerylib_0.1.4    
#> [43] MASS_7.3-66         cachem_1.1.0        abind_1.4-8        
#> [46] mime_0.13           ggstats_0.14.0      tidyselect_1.2.1   
#> [49] digest_0.6.39       mvtnorm_1.4-2       purrr_1.2.2        
#> [52] dplyr_1.2.1         labeling_0.4.3      forcats_1.0.1      
#> [55] labelled_2.16.1     fastmap_1.2.0       grid_4.6.2         
#> [58] cli_3.6.6           magrittr_2.0.5      broom_1.0.13       
#> [61] withr_3.0.3         backports_1.5.1     FPDclustering_2.3.6
#> [64] scales_1.4.0        promises_1.5.0      rmarkdown_2.32     
#> [67] otel_0.3.0          ggsignif_0.6.4      ragg_1.5.2         
#> [70] hms_1.1.4           shiny_1.14.0        evaluate_1.0.5     
#> [73] knitr_1.52          haven_2.5.5         ggeasy_0.1.6       
#> [76] miniUI_0.1.2        rlang_1.3.0         Rcpp_1.1.2         
#> [79] xtable_1.8-8        glue_1.8.1          ExPosition_2.11.0  
#> [82] ThreeWay_1.1.4      rstudioapi_0.19.0   jsonlite_2.0.0     
#> [85] questionr_0.8.2     R6_2.6.1            systemfonts_1.3.2  
#> [88] fs_2.1.0

References

Bezdek, James C. 1981. Pattern Recognition with Fuzzy Objective Function Algorithms. Springer. https://doi.org/10.1007/978-1-4757-0450-1.
Campello, Ricardo J. G. B., and Eduardo R. Hruschka. 2006. “A Fuzzy Extension of the Silhouette Width Criterion for Cluster Analysis.” Fuzzy Sets and Systems 157 (21): 2858–75. https://doi.org/10.1016/j.fss.2006.07.006.
Iyigun, Cem, and Adi Ben-Israel. 2008. “Probabilistic Distance Clustering Adjusted for Cluster Size.” Probability in the Engineering and Informational Sciences 22 (4): 603–21. https://doi.org/10.1017/S0269964808000351.
Menardi, Giovanna. 2011. “Density-Based Silhouette Diagnostics for Clustering Methods.” Statistics and Computing 21 (3): 295–308. https://doi.org/10.1007/s11222-010-9169-0.
Scrucca, Luca, Michael Fop, T. Brendan Murphy, and Adrian E. Raftery. 2016. “mclust 5: Clustering, Classification and Density Estimation Using Gaussian Finite Mixture Models.” The R Journal 8 (1): 289–317. https://doi.org/10.32614/RJ-2016-021.
Tortora, Cristina, and Francesco Palumbo. 2025. “FPDclustering: A Comprehensive R Package for Probabilistic Distance Clustering Based Methods.” Computational Statistics 40 (2): 1123–46. https://doi.org/10.1007/s00180-024-01490-5.