| Type: | Package |
| Title: | Proximity Measure Based Diagnostics for Standard, Soft, and Multi-Way Clustering |
| Version: | 0.9.7 |
| Depends: | R (≥ 4.2.0) |
| Language: | en-US |
| Maintainer: | Shrikrishna Bhat K <skbhat.in@gmail.com> |
| Description: | Quantifies clustering quality by measuring both cohesion within clusters and separation between clusters. Silhouette widths are computed from the n by K matrices that clustering algorithms return (distances or similarities to cluster prototypes, fuzzy memberships, or posterior probabilities) instead of all pairwise dissimilarities, for the partition an algorithm produced or one supplied by the user. Implements the simplified silhouette of Van der Laan et al. (2003) <doi:10.1080/0094965031000136012>, the probability of alternative cluster normalization of Raymaekers and Rousseeuw (2022) <doi:10.1080/10618600.2022.2050249>, the fuzzy silhouette average of Campello and Hruschka (2006) <doi:10.1016/j.fss.2006.07.006>, the density-based silhouette of Menardi (2011) <doi:10.1007/s11222-010-9169-0>, the soft silhouettes of Bhat Kapu and Kiruthika (2024) <doi:10.1080/23737484.2024.2408534>, and the extended silhouette for multi-way clustering of Schepers et al. (2008) <doi:10.1007/s00357-008-9005-9> and Bhat Kapu and Kiruthika (2025) <doi:10.21203/rs.3.rs-6973596/v1>. Provides summaries and 'ggplot2' silhouette plots in the style of Rousseeuw (1987) <doi:10.1016/0377-0427(87)90125-7>. |
| License: | GPL-2 |
| URL: | https://kskbhat.github.io/Silhouette/ |
| BugReports: | https://github.com/kskbhat/Silhouette/issues |
| Encoding: | UTF-8 |
| Imports: | ggplot2, ggpubr, methods, stats, utils |
| Suggests: | proxy, ppclust, cluster, factoextra, drclust, mclust, knitr, lifecycle, rmarkdown, spelling, testthat (≥ 3.0.0) |
| VignetteBuilder: | knitr |
| Config/Needs/website: | FPDclustering, bioc::bluster |
| Config/testthat/edition: | 3 |
| NeedsCompilation: | no |
| Config/roxygen2/markdown: | TRUE |
| Config/roxygen2/version: | 8.0.0 |
| Packaged: | 2026-11-15 15:42:16 UTC; root |
| Author: | Shrikrishna Bhat K
|
| Repository: | CRAN |
| Date/Publication: | 2026-11-15 21:30:17 UTC |
Silhouette: Proximity Measure Based Diagnostics for Standard, Soft, and Multi-Way Clustering
Description
Quantifies clustering quality by measuring both cohesion within clusters and separation between clusters. Silhouette widths are computed from the n by K matrices that clustering algorithms return (distances or similarities to cluster prototypes, fuzzy memberships, or posterior probabilities) instead of all pairwise dissimilarities, for the partition an algorithm produced or one supplied by the user. Implements the simplified silhouette of Van der Laan et al. (2003) doi:10.1080/0094965031000136012, the probability of alternative cluster normalization of Raymaekers and Rousseeuw (2022) doi:10.1080/10618600.2022.2050249, the fuzzy silhouette average of Campello and Hruschka (2006) doi:10.1016/j.fss.2006.07.006, the density-based silhouette of Menardi (2011) doi:10.1007/s11222-010-9169-0, the soft silhouettes of Bhat Kapu and Kiruthika (2024) doi:10.1080/23737484.2024.2408534, and the extended silhouette for multi-way clustering of Schepers et al. (2008) doi:10.1007/s00357-008-9005-9 and Bhat Kapu and Kiruthika (2025) doi:10.21203/rs.3.rs-6973596/v1. Provides summaries and 'ggplot2' silhouette plots in the style of Rousseeuw (1987) doi:10.1016/0377-0427(87)90125-7.
Author(s)
Maintainer: Shrikrishna Bhat K skbhat.in@gmail.com (ORCID) [copyright holder]
Authors:
Shrikrishna Bhat K skbhat.in@gmail.com (ORCID) [copyright holder]
Kiruthika C (ORCID)
See Also
Useful links:
Calculate Silhouette Widths, Summary, and Plot for Clustering Results
Description
Computes the simplified silhouette width of each observation from an
n \times K matrix of proximities between observations and clusters
(for example, distances to cluster centers). The width compares the
proximity of an observation to its own cluster with its proximity to the
closest other cluster. Widths lie in [-1, 1]; values near 1 indicate a
clear assignment and negative values indicate that another cluster is closer.
Usage
Silhouette(
prox_matrix,
proximity_type = c("dissimilarity", "similarity"),
method = c("medoid", "pac"),
average = c("crisp", "fuzzy", "median"),
prob_matrix = NULL,
a = 2,
sort = FALSE,
print.summary = FALSE,
clust_fun = NULL,
...,
cluster = NULL,
clust_args = list()
)
## S3 method for class 'Silhouette'
plot(
x,
label = FALSE,
summary.legend = TRUE,
grayscale = FALSE,
linetype = c("dashed", "solid", "dotted", "dotdash", "longdash", "twodash"),
...
)
## S3 method for class 'Silhouette'
summary(object, print.summary = TRUE, ...)
Arguments
prox_matrix |
A numeric matrix where rows represent observations and columns represent proximity measures (e.g., distances or similarities) to clusters. Typically, this is a membership or dissimilarity matrix from clustering results. If |
proximity_type |
Character string specifying the type of proximity measure in |
method |
Character string specifying the normalization of the silhouette width. Options are |
average |
Character string specifying the method for computing the average silhouette width. Options are:
Defaults to |
prob_matrix |
A numeric matrix of cluster membership probabilities, where rows represent
observations and columns represent clusters in the same order as |
a |
Non-negative number, the exponent |
sort |
Logical; if |
print.summary |
Logical; if |
clust_fun |
Optional S3 or S4 function object or function as character string specifying a clustering function that produces the proximity measure matrix. For example, |
... |
Additional arguments passed to |
cluster |
Optional integer vector of length |
clust_args |
A list of further arguments for |
x |
An object of class |
label |
Logical; if |
summary.legend |
Logical; if |
grayscale |
Logical; if |
linetype |
Character or numeric value specifying the type of line to be used for the horizontal reference line indicating the average silhouette width. Accepts standard ggplot2 linetype values, such as:
Defaults to |
object |
An object of class |
Details
Silhouette() is not the same as silhouette from
the cluster package: that function needs all n(n-1)/2 pairwise
dissimilarities, whereas Silhouette() needs only the n \times K
proximities between observations and cluster representatives (centroids,
medoids or prototypes). The two give different, although usually similar,
values.
The Silhouette function implements the Simplified Silhouette method introduced by Van der Laan, Pollard, & Bryan (2003), which adapts and generalizes the classic silhouette method of Rousseeuw (1987).
Clustering quality is evaluated using a proximity matrix, denoted as
\Delta = [\delta_{ik}]_{n \times K} for dissimilarity measures or
\Delta' = [\delta'_{ik}]_{n \times K} for similarity measures.
Here, i = 1, \ldots, n indexes observations, and k = 1, \ldots, K indexes clusters.
\delta_{ik} represents the dissimilarity (e.g., distance) between observation i and cluster k,
while \delta'_{ik} represents similarity values.
Let c_i be the cluster of observation i (see Cluster labels below)
and \nu_i its neighboring cluster, the closest cluster other than c_i.
The silhouette width S(x_i) for observation i depends on the proximity type:
For dissimilarity measures:
S(x_i) = \frac{ \delta_{i\nu_i} - \delta_{ic_i} }{ N(x_i) }
For similarity measures:
S(x_i) = \frac{ \delta'_{ic_i} - \delta'_{i\nu_i} }{ N(x_i) }
where N(x_i) is a normalizing factor defined by the method.
Choice of method:
The normalizer N(x_i) is selected according to the method argument:
For
medoid(Van der Laan et al., 2003):Dissimilarity:
\max(\delta_{ic_i}, \delta_{i\nu_i})Similarity:
\max(\delta'_{ic_i}, \delta'_{i\nu_i})
For
pac(Raymaekers & Rousseeuw, 2022):Dissimilarity:
\delta_{ic_i} + \delta_{i\nu_i}Similarity:
\delta'_{ic_i} + \delta'_{i\nu_i}
The names describe the normalization, not the algorithm. The name "medoid" is kept
for compatibility with earlier versions and with Van der Laan et al. (2003), who
introduced this width for partitioning around medoids; it divides by the larger of the two
proximities and applies equally to centroids (as in k-means), prototypes, or any other
cluster representative. Likewise, "pac" does not require a classifier. For the same
labels and non-negative proximities, the PAC width equals s/(2 - |s|), where s
is the medoid width, so the two rank observations identically and differ only in scale.
These methods can also be applied to proximities derived from classification algorithms, to assess the separation of groups formed by a classifier.
Two equal proximities give a width of 0 (including two zero distances). An infinite proximity to the neighboring cluster gives a width of 1 and an infinite proximity to the assigned cluster gives -1.
Value
A data frame of class "Silhouette" with one row per observation and columns
cluster, neighbor, weight (only when prob_matrix is supplied) and
sil_width. The object includes the following attributes:
- proximity_type
The proximity type used (
"similarity"or"dissimilarity").- method
The silhouette calculation method used (
"medoid"or"pac").- average
Character — the averaging method:
"crisp","fuzzy", or"median".
Further, summary returns (invisibly) a list containing:
-
clus.avg.widths: A named numeric vector of average silhouette widths per cluster. -
avg.width: The overall average silhouette width. -
sil.sum: A data frame with columnscluster,size, andavg.sil.widthsummarizing cluster sizes and average silhouette widths.
Cluster labels
The cluster c_i of each observation is taken, in order of precedence, from
-
cluster, when supplied; the largest membership in
prob_matrix, when supplied;the closest cluster in
prox_matrix.
The neighboring cluster \nu_i is always the closest cluster other than c_i in
the proximity matrix. When the labels come from the proximity matrix itself (case 3), every
width is non-negative, so a poorly assigned observation shows up only as a width near 0.
When the labels come from elsewhere (cases 1 and 2), an observation that is closer to another
cluster than to its own receives a negative width. To assess the partition produced by a
clustering algorithm, supply that partition through cluster or prob_matrix.
Averages
If average = "crisp", the crisp silhouette index (CS) is the mean width:
CS = \frac{1}{n} \sum_{i=1}^{n} S(x_i)
summarizing overall clustering quality.
If average = "fuzzy" and prob_matrix is provided, denoted as \Gamma = [\gamma_{ik}]_{n \times K},
with \gamma_{ik} representing the probability of observation i belonging to cluster k,
the fuzzy silhouette index (FS) of Campello and Hruschka (2006) is:
FS = \frac{\sum_{i=1}^{n} w_i S(x_i) }{\sum_{i=1}^{n} w_i},
\qquad w_i = ( \gamma_{i(1)} - \gamma_{i(2)} )^{\alpha},
where \gamma_{i(1)} and \gamma_{i(2)} are the largest and second largest
memberships of observation i, and \alpha (the a argument) controls the
emphasis on confident assignments (\alpha \ge 0; \alpha = 0 gives the crisp
average). The weight w_i is stored in the weight column and does not depend on
the cluster labels; it equals Equation (8) of Bhat Kapu and Kiruthika (2024) when each
observation is assigned to its largest membership. If all weights are zero, the fuzzy
average is undefined (NaN) and a warning is given.
If average = "median", the median width is reported.
Missing widths are removed before averaging.
Input checks and ties
prox_matrix must be a numeric matrix with at least two columns and no missing
values; +Inf is allowed (for example, -\log 0) but -Inf is not.
Negative proximities are rejected, because the widths would then no longer lie in
[-1, 1]; transform them first (for example, (1 + r)/2 for correlations).
prob_matrix must be non-negative with rows summing to 1 within a tolerance of
sqrt(.Machine$double.eps) (about 1.5 \times 10^{-8}); rows are not
renormalized.
The closest cluster, the neighboring cluster and the largest membership are found with
max.col(..., ties.method = "first"). Exactly tied proximities or memberships
are therefore resolved in favor of the column that comes first, so with exact ties the
labels (though not the widths from the proximity matrix itself, which are then 0) depend on
the order of the columns. Supply cluster to control the assignment of tied
observations.
References
Rousseeuw, P. J. (1987). Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20, 53–65. doi:10.1016/0377-0427(87)90125-7
Van der Laan, M., Pollard, K., & Bryan, J. (2003). A new partitioning around medoids algorithm. Journal of Statistical Computation and Simulation, 73(8), 575–584. doi:10.1080/0094965031000136012
Campello, R. J., & Hruschka, E. R. (2006). A fuzzy extension of the silhouette width criterion for cluster analysis. Fuzzy Sets and Systems, 157(21), 2858–2875. doi:10.1016/j.fss.2006.07.006
Raymaekers, J., & Rousseeuw, P. J. (2022). Silhouettes and quasi residual plots for neural nets and tree-based classifiers. Journal of Computational and Graphical Statistics, 31(4), 1332–1343. doi:10.1080/10618600.2022.2050249
Bhat Kapu, S., & Kiruthika. (2024). Some density-based silhouette diagnostics for soft clustering algorithms. Communications in Statistics: Case Studies, Data Analysis and Applications, 10(3-4), 221-238. doi:10.1080/23737484.2024.2408534
See Also
softSilhouette, dbSilhouette, cerSilhouette, calSilhouette, getSilhouette, is.Silhouette, plotSilhouette, silhouette
Examples
# Simplified silhouette of a k-means partition of the iris data
data(iris)
out <- kmeans(iris[, -5], 3)
if (requireNamespace("proxy", quietly = TRUE)) {
d <- proxy::dist(iris[, -5], out$centers)
silh_out <- Silhouette(d, print.summary = TRUE)
plot(silh_out)
# Score a partition produced elsewhere: average-linkage clustering,
# with the distances to the centroids of its groups
hc <- cutree(hclust(dist(iris[, -5]), "average"), 3)
centers <- apply(iris[, -5], 2, function(v) tapply(v, hc, mean))
silh_hc <- Silhouette(proxy::dist(iris[, -5], centers), cluster = hc)
summary(silh_hc)
} else {
message("Install 'proxy': install.packages('proxy')")
}
# Scree plot for optimal clusters (2 to 7)
if (requireNamespace("ppclust", quietly = TRUE)) {
library(ppclust)
avg_sil_width <- rep(NA,7)
for (k in 2:7) {
out <- Silhouette(
prox_matrix = "d",
proximity_type = "dissimilarity",
prob_matrix = "u",
clust_fun = ppclust::fcm,
x = iris[, 1:4],
centers = k,
average = "fuzzy"
)
# Compute average silhouette width from widths
avg_sil_width[k] <- summary(out, print.summary = FALSE)$avg.width
}
plot(avg_sil_width,
type = "o",
ylab = "Overall Silhouette Width",
xlab = "Number of Clusters",
main = "Scree Plot"
)
} else {
message("Install 'ppclust': install.packages('ppclust')")
}
Calculate All Available Silhouette Summaries
Description
Computes every silhouette width available in the package for one clustering and returns a data frame comparing the crisp, fuzzy, and median averages of each.
Usage
calSilhouette(
prox_matrix = NULL,
proximity_type = c("dissimilarity", "similarity"),
prob_matrix = NULL,
a = 2,
print.summary = FALSE,
clust_fun = NULL,
...,
cluster = NULL,
clust_args = list()
)
Arguments
prox_matrix |
A numeric matrix where rows represent observations and columns represent proximity measures (e.g., distances or similarities) to clusters. Typically, this is a membership or dissimilarity matrix from clustering results. If |
proximity_type |
Character string specifying the type of proximity measure in |
prob_matrix |
A numeric matrix of cluster membership probabilities, where rows represent
observations and columns represent clusters in the same order as |
a |
Non-negative number, the exponent |
print.summary |
Logical; if |
clust_fun |
Optional S3 or S4 function object or function as character string specifying a clustering function that produces the proximity measure matrix. For example, |
... |
Additional arguments passed to |
cluster |
Optional integer vector of length |
clust_args |
A list of further arguments for |
Details
The rows are not independent pieces of evidence: several are exact or monotone transformations of one another (see Relations between the rows). The distance-based rows compare distances to the cluster representatives; the probability-based rows are the silhouettes of Bhat Kapu and Kiruthika (2024), which treat the posterior probabilities or the component densities as proximities, and the density-based silhouette of Menardi (2011).
The methods included depend on the available input matrices:
If prox_matrix is available:
-
medoid- Medoid-based silhouette usingSilhouette -
pac- PAC-based silhouette usingSilhouette
If prob_matrix is available:
-
pp_pac- PACS: posterior probabilities with the PAC normalization (softSilhouette) -
pp_medoid- PPS: posterior probability silhouette (softSilhouette) -
nlpp_pac- Negative log posterior probabilities with the PAC normalization (softSilhouette) -
nlpp_medoid- NLPPS: negative log of posterior probability silhouette (softSilhouette) -
pd_pac- Probability densities with the PAC normalization (softSilhouette) -
pd_medoid- PDS: probability density silhouette (softSilhouette) -
cer- CeS: certainty silhouette (cerSilhouette) -
db- DBS: density-based silhouette of Menardi (2011) (dbSilhouette)
At least one of prox_matrix or prob_matrix must be provided. Each width is
computed once and summarized by all available averages; the fuzzy average needs
prob_matrix.
If a width cannot be computed, its row is NA and a warning gives the reason.
Warnings from the individual functions are collected and each distinct warning is
reported once.
Value
A data frame with one row per method and the following columns:
- Method
Character vector of method names, as listed in Details.
- Crisp
Crisp (unweighted mean) average silhouette width.
- Fuzzy
Fuzzy (membership-weighted) average silhouette width; present only when
prob_matrixis supplied.- Median
Median silhouette width.
Cluster labels
When cluster is supplied, every row scores that partition. Otherwise every row
except the PDS rows (pd_pac, pd_medoid) uses the largest membership in
prob_matrix, or the closest cluster in prox_matrix when there are no
memberships. The PDS rows assign each observation to its largest density, as in the
definition of the PDS (see softSilhouette); this agrees with the largest
membership when the clusters have equal estimated weights. (In version 0.9.6 and earlier,
the crisp and median distance-based rows used the closest cluster in prox_matrix
even when prob_matrix was supplied.)
Relations between the rows
For the same labels, the rows fall into the following groups.
- Monotone transformations of each other (same ranking of observations).
-
Each
pacwidth iss/(2 - |s|)of the correspondingmedoidwidths:medoid/pac,pp_medoid/pp_pac,nlpp_medoid/nlpp_pacandpd_medoid/pd_pac. With the default labels,pp_medoid,pp_pacanddbare all increasing functions of the ratio of the two largest memberships, so they too rank observations identically. The averages still differ, and may peak at different numbers of clusters. - Equal in special cases.
-
pd(PDS) equalspp(PPS) when all estimated cluster weights are equal. For fuzzy c-means with fuzzifierm,pp_medoidandpp_pacequalmedoidandpaccomputed from the dissimilaritiesd_{ik}^{1/(m-1)}, whered_{ik}are squared Euclidean distances to the prototypes (form = 2, thedcomponent offcm). - Genuinely different.
-
The distance-based rows (in general),
nlpp(which also depends on the size of the neighboring membership), andcer(the largest membership alone, bounded below by1/K).
Choosing the number of clusters
Computing calSilhouette() over k = 2, 3, \ldots gives one curve per row.
The distance-based rows are usually read at their maximum. Bhat Kapu and Kiruthika (2024)
read the probability-based rows from the reverse elbow drop of the curve, and the DBS from
its local maxima (Menardi, 2011). The suggested number of clusters depends on the
algorithm that produced the memberships, on the width and on the averaging.
References
Bhat Kapu, S., & Kiruthika. (2024). Some density-based silhouette diagnostics for soft clustering algorithms. Communications in Statistics: Case Studies, Data Analysis and Applications, 10(3-4), 221-238. doi:10.1080/23737484.2024.2408534
Menardi, G. (2011). Density-based silhouette diagnostics for clustering methods. Statistics and Computing, 21(3), 295–308. doi:10.1007/s11222-010-9169-0
See Also
Silhouette, softSilhouette, dbSilhouette, cerSilhouette
Examples
if (requireNamespace("ppclust", quietly = TRUE)) {
# Example with FCM clustering
library(ppclust)
data(iris)
fcm_result <- fcm(iris[, -5], centers = 3)
# Using matrices directly
summary_result <- calSilhouette(
prox_matrix = fcm_result$d,
prob_matrix = fcm_result$u,
proximity_type = "dissimilarity",
print.summary = TRUE
)
}
if (requireNamespace("ppclust", quietly = TRUE)) {
# Using clustering function
summary_result2 <- calSilhouette(
prox_matrix = "d",
prob_matrix = "u",
proximity_type = "dissimilarity",
clust_fun = ppclust::fcm,
x = iris[, -5],
centers = 3,
print.summary = TRUE
)
}
Certainty Silhouette Width (CeS) for Soft Clustering
Description
Computes the certainty silhouette (CeS) of Bhat Kapu and Kiruthika (2024) for each observation: the membership probability of its assigned cluster (by default, its largest membership).
Usage
cerSilhouette(
prob_matrix,
average = c("crisp", "fuzzy", "median"),
a = 2,
sort = FALSE,
print.summary = FALSE,
clust_fun = NULL,
...,
cluster = NULL,
clust_args = list()
)
Arguments
prob_matrix |
A numeric matrix of posterior probabilities where rows represent observations and columns represent clusters. Must sum to 1 by row. If |
average |
Character string specifying the method for computing the average silhouette width. Options are:
Defaults to |
a |
Non-negative number, the exponent |
sort |
Logical; if |
print.summary |
Logical; if |
clust_fun |
Optional S3 or S4 function object or function as character string specifying a clustering function that produces the membership matrix. For example, |
... |
Additional arguments passed to |
cluster |
Optional integer vector of length |
clust_args |
A list of further arguments for |
Details
The certainty silhouette is the counterpart of the uncertainty used in model-based
clustering: values near 1 indicate high certainty that an observation belongs to its
cluster, and values near 1/K ambiguity. It is bounded below by 1/K, so its
level falls as K grows; Bhat Kapu and Kiruthika (2024) read the number of clusters
from the reverse elbow drop of its average over k (see softSilhouette).
Let the posterior probability matrix or cluster membership matrix be
\Gamma = [\gamma_{ik}]_{n \times K}.
The certainty silhouette width for observation i assigned to cluster c_i is:
\mathrm{Cer}_i = \gamma_{ic_i},
which equals \max_{k} \gamma_{ik} \in [1/K, 1] with the default labels. With labels
supplied through cluster, the width lies in [0, 1].
The neighbor column gives the cluster with the largest membership other than
c_i.
Ties are resolved in favor of the first column; see Silhouette.
Value
A data frame of class "Silhouette" with one row per observation and columns
cluster, neighbor, weight (only when average = "fuzzy") and
sil_width. The object includes the following attributes:
- proximity_type
The proximity type used i.e.,
"similarity".- method
The silhouette calculation method used i.e.,
"certainty".- average
Character — the averaging method:
"crisp","fuzzy", or"median".
Averages
If average = "crisp", the crisp silhouette index (CS) is the mean width:
CS = \frac{1}{n} \sum_{i=1}^{n} S(x_i)
summarizing overall clustering quality.
If average = "fuzzy" and prob_matrix is provided, denoted as \Gamma = [\gamma_{ik}]_{n \times K},
with \gamma_{ik} representing the probability of observation i belonging to cluster k,
the fuzzy silhouette index (FS) of Campello and Hruschka (2006) is:
FS = \frac{\sum_{i=1}^{n} w_i S(x_i) }{\sum_{i=1}^{n} w_i},
\qquad w_i = ( \gamma_{i(1)} - \gamma_{i(2)} )^{\alpha},
where \gamma_{i(1)} and \gamma_{i(2)} are the largest and second largest
memberships of observation i, and \alpha (the a argument) controls the
emphasis on confident assignments (\alpha \ge 0; \alpha = 0 gives the crisp
average). The weight w_i is stored in the weight column and does not depend on
the cluster labels; it equals Equation (8) of Bhat Kapu and Kiruthika (2024) when each
observation is assigned to its largest membership. If all weights are zero, the fuzzy
average is undefined (NaN) and a warning is given.
If average = "median", the median width is reported.
Missing widths are removed before averaging.
References
Bhat Kapu, S., & Kiruthika. (2024). Some density-based silhouette diagnostics for soft clustering algorithms. Communications in Statistics: Case Studies, Data Analysis and Applications, 10(3-4), 221-238. doi:10.1080/23737484.2024.2408534
See Also
Silhouette, softSilhouette, dbSilhouette, calSilhouette, getSilhouette, is.Silhouette, plotSilhouette
Examples
prob <- matrix(c(0.90, 0.10,
0.60, 0.40,
0.20, 0.80,
0.45, 0.55), ncol = 2, byrow = TRUE)
cerSilhouette(prob)
# Compare two soft clustering algorithms on the iris data
data(iris)
if (requireNamespace("ppclust", quietly = TRUE)) {
fcm_result <- ppclust::fcm(iris[, 1:4], 3)
out_fcm <- cerSilhouette(prob_matrix = fcm_result$u, print.summary = TRUE)
plot(out_fcm)
fcm2_result <- ppclust::fcm2(iris[, 1:4], 3)
out_fcm2 <- cerSilhouette(prob_matrix = fcm2_result$u)
c(fcm = summary(out_fcm, print.summary = FALSE)$avg.width,
fcm2 = summary(out_fcm2, print.summary = FALSE)$avg.width)
} else {
message("Install 'ppclust' to run this example: install.packages('ppclust')")
}
Density-Based Silhouette Width (DBS) for Soft Clustering
Description
Computes the density-based silhouette (DBS) of Menardi (2011), using log-ratios of posterior probabilities.
Usage
dbSilhouette(
prob_matrix,
average = c("median", "crisp", "fuzzy"),
a = 2,
sort = FALSE,
print.summary = FALSE,
clust_fun = NULL,
...,
cluster = NULL,
clust_args = list()
)
Arguments
prob_matrix |
A numeric matrix of posterior probabilities where rows represent observations and columns represent clusters. Must sum to 1 by row. If |
average |
Character string specifying the method for computing the average silhouette width. Options are:
Defaults to |
a |
Non-negative number, the exponent |
sort |
Logical; if |
print.summary |
Logical; if |
clust_fun |
Optional S3 or S4 function object or function as character string specifying a clustering function that produces the membership matrix. For example, |
... |
Additional arguments passed to |
cluster |
Optional integer vector of length |
clust_args |
A list of further arguments for |
Details
The DBS compares the largest membership with the second largest on the log scale, scaled
by the largest such log ratio in the sample. Its numerator is that of the NLPPS of Bhat
Kapu and Kiruthika (2024), with a different denominator. Following Menardi (2011), the
median is the default summary, and the number of clusters is read from the local maxima
of the average over k.
Let the posterior probability matrix or cluster membership matrix be
\Gamma = [\gamma_{ik}]_{n \times K}, let c_i be the cluster of observation
i and \nu_i the cluster with its largest membership other than c_i.
With the log ratio
R_i = \log \frac{\gamma_{ic_i}}{\gamma_{i\nu_i}},
the density-based silhouette width is
\mathrm{DBS}_i = \frac{R_i}{\max_{j=1,\dots,n} |R_j|}.
With the default labels (largest membership) R_i \ge 0, so the widths lie in
[0, 1] and the most confidently assigned observation has width 1. With labels
supplied through cluster, R_i can be negative and the widths lie in
[-1, 1].
Because of the scaling, the average width depends on how extreme the most confident observation is: a single observation with a very large log ratio shrinks every other width towards zero. Compare the average only across fits with similar extremes, and report it together with the mode-wise or cluster-wise summaries.
Value
A data frame of class "Silhouette" with one row per observation and columns
cluster, neighbor, weight (only when average = "fuzzy") and
sil_width. The object includes the following attributes:
- proximity_type
The proximity type used i.e.,
"similarity".- method
The silhouette calculation method used i.e.,
"db".- average
Character — the averaging method:
"crisp","fuzzy", or"median".
Numerical safeguards
The following rules keep the widths finite and meaningful:
-
Zero probabilities. When the neighboring membership is exactly zero (for example, after a posterior probability underflows),
R_i = +\infty. Such observations receive width 1 (and width -1 when the assigned membership is zero, which can happen only with supplied labels), and a warning reports how many there are. The remaining widths are scaled by the largest finite|R_j|. (In version 0.9.6 and earlier, the infinite maximum turned every finite width into 0.) -
Equal memberships. When the two memberships are equal,
R_i = 0and the width is 0. -
Degenerate fits. When the largest finite
|R_j|is belowsqrt(.Machine$double.eps)(about1.5 \times 10^{-8}), the memberships are numerically indistinguishable and scaling would only magnify rounding error. The finite widths are then set to 0 with a warning.
Ties are resolved in favor of the first column; see Silhouette.
Averages
If average = "crisp", the crisp silhouette index (CS) is the mean width:
CS = \frac{1}{n} \sum_{i=1}^{n} S(x_i)
summarizing overall clustering quality.
If average = "fuzzy" and prob_matrix is provided, denoted as \Gamma = [\gamma_{ik}]_{n \times K},
with \gamma_{ik} representing the probability of observation i belonging to cluster k,
the fuzzy silhouette index (FS) of Campello and Hruschka (2006) is:
FS = \frac{\sum_{i=1}^{n} w_i S(x_i) }{\sum_{i=1}^{n} w_i},
\qquad w_i = ( \gamma_{i(1)} - \gamma_{i(2)} )^{\alpha},
where \gamma_{i(1)} and \gamma_{i(2)} are the largest and second largest
memberships of observation i, and \alpha (the a argument) controls the
emphasis on confident assignments (\alpha \ge 0; \alpha = 0 gives the crisp
average). The weight w_i is stored in the weight column and does not depend on
the cluster labels; it equals Equation (8) of Bhat Kapu and Kiruthika (2024) when each
observation is assigned to its largest membership. If all weights are zero, the fuzzy
average is undefined (NaN) and a warning is given.
If average = "median", the median width is reported.
Missing widths are removed before averaging.
References
Menardi, G. (2011). Density-based silhouette diagnostics for clustering methods. Statistics and Computing, 21(3), 295–308. doi:10.1007/s11222-010-9169-0
Bhat Kapu, S., & Kiruthika. (2024). Some density-based silhouette diagnostics for soft clustering algorithms. Communications in Statistics: Case Studies, Data Analysis and Applications, 10(3-4), 221-238. doi:10.1080/23737484.2024.2408534
See Also
Silhouette, softSilhouette, cerSilhouette, calSilhouette, getSilhouette, is.Silhouette, plotSilhouette
Examples
prob <- matrix(c(0.90, 0.10,
0.60, 0.40,
0.20, 0.80,
0.45, 0.55), ncol = 2, byrow = TRUE)
dbSilhouette(prob)
# A zero neighboring probability gives width 1 and a warning
prob0 <- rbind(prob, c(1, 0))
dbSilhouette(prob0)
# Compare two soft clustering algorithms on the iris data
data(iris)
if (requireNamespace("ppclust", quietly = TRUE)) {
fcm_result <- ppclust::fcm(iris[, 1:4], 3)
out_fcm <- dbSilhouette(prob_matrix = fcm_result$u, print.summary = TRUE)
plot(out_fcm)
fcm2_result <- ppclust::fcm2(iris[, 1:4], 3)
out_fcm2 <- dbSilhouette(prob_matrix = fcm2_result$u)
c(fcm = summary(out_fcm, print.summary = FALSE)$avg.width,
fcm2 = summary(out_fcm2, print.summary = FALSE)$avg.width)
} else {
message("Install 'ppclust' to run this example: install.packages('ppclust')")
}
Calculate Extended Silhouette Width for Multi-Way Clustering
Description
Computes an extended silhouette width for multi-way clustering (e.g., biclustering, triclustering, or n-mode tensor clustering) by combining silhouette widths from a list of Silhouette objects, each representing one mode of clustering. The extended silhouette width is a weighted average of the average silhouette widths of the modes. By default each mode is weighted by its number of objects; equal or user-specified weights are also available. The output is an object of class extSilhouette.
Usage
extSilhouette(
sil_list,
dim_names = NULL,
print.summary = FALSE,
weights = c("size", "equal")
)
## S3 method for class 'extSilhouette'
format(x, digits = 4, ...)
## S3 method for class 'extSilhouette'
print(x, digits = 4, ...)
Arguments
sil_list |
A list of objects of class |
dim_names |
An optional character vector of dimension names (e.g., |
print.summary |
Logical; if |
weights |
How the modes are weighted: |
x |
An object of class |
digits |
Number of decimal places shown. |
... |
Further arguments, currently ignored. |
Details
The extended silhouette width is computed as:
ExS = \frac{ \sum_{i} \omega_i \bar{S}_i }{ \sum_{i} \omega_i }
where \bar{S}_i is the average silhouette width of mode i (from
summary(x)$avg.width, so it follows the averaging rule of that object) and
\omega_i is its weight. With weights = "size", \omega_i = n_i, the
number of objects in mode i (nrow(x)), as in Schepers et al. (2008); ExS is
then the average width over all objects of all modes, each object counting once. With
weights = "equal", \omega_i = 1, so ExS is the mean of the mode-wise
averages.
Size weighting treats every clustered object alike, but when the modes differ greatly in
size the largest mode dominates: with 200 rows and 30 columns, the rows receive about 87\
of the weight. Equal weighting gives every mode the same say regardless of its size.
Because the choice can change which model is selected, report the mode-wise averages in
dim_table together with ExS.
Each Silhouette object in sil_list must contain at least one observation and have a finite average width. For consistency, compute all Silhouette objects with the same function, method and averaging rule.
Value
A list of class "extSilhouette" with the following components:
- ext_sil_width
A numeric scalar, the extended silhouette width.
- dim_table
A data frame with columns
dimension(e.g., "Mode 1", "Mode 2"),n_obs(number of objects),weight(the share of the total weight given to the mode) andavg_sil_width(average silhouette width of the mode).- weights
The weighting used:
"size","equal"or"user".
Values are not rounded; the print method shows four decimals.
References
Schepers, J., Ceulemans, E., & Van Mechelen, I. (2008). Selecting among multi-mode partitioning models of different complexities: A comparison of four model selection criteria. Journal of Classification, 25(1), 67–85. doi:10.1007/s00357-008-9005-9
Bhat Kapu, S., & Kiruthika, C. (2025). Block Probabilistic Distance Clustering: A Unified Framework and Evaluation. PREPRINT (Version 1) available at Research Square. doi:10.21203/rs.3.rs-6973596/v1
See Also
Silhouette, softSilhouette, dbSilhouette, cerSilhouette, getSilhouette, is.Silhouette
Examples
# Two modes: 150 objects in 3 clusters and 4 variables in 2 clusters
set.seed(123)
prob_mode1 <- matrix(runif(150 * 3), nrow = 150, ncol = 3)
prob_mode1 <- prob_mode1 / rowSums(prob_mode1)
prob_mode2 <- matrix(runif(4 * 2), nrow = 4, ncol = 2)
prob_mode2 <- prob_mode2 / rowSums(prob_mode2)
sil_mode1 <- softSilhouette(prob_matrix = prob_mode1, method = "pac")
sil_mode2 <- softSilhouette(prob_matrix = prob_mode2, method = "pac")
# Size-weighted (default) and equally weighted extended silhouette
extSilhouette(list(sil_mode1, sil_mode2), dim_names = c("Rows", "Columns"))
extSilhouette(list(sil_mode1, sil_mode2), dim_names = c("Rows", "Columns"),
weights = "equal")
Create Silhouette Object from User Components
Description
Constructs a Silhouette class object directly from user-provided components without performing silhouette calculations. This function allows users to build a Silhouette object when they already have the necessary components.
Usage
getSilhouette(
cluster,
neighbor,
sil_width,
weight = NULL,
proximity_type = c("dissimilarity", "similarity"),
method = NA,
average = c("crisp", "fuzzy", "median")
)
Arguments
cluster |
Numeric or integer vector of cluster assignments for each observation |
neighbor |
Numeric or integer vector of nearest neighbor cluster assignments for each observation |
sil_width |
Numeric vector of silhouette widths for each observation (must be between -1 and +1) |
weight |
Numeric vector of weights for each observation (must be between 0 and 1, only used when average = "fuzzy") |
proximity_type |
Character; the proximity type used. Options: "similarity" or "dissimilarity" |
method |
Character; the silhouette calculation method used (default: |
average |
Character; the averaging method. Options: "crisp", "fuzzy", or "median" |
Value
A data frame of class "Silhouette" containing cluster assignments, nearest neighbor clusters, silhouette widths for each observation, and weights (for fuzzy clustering). The object includes the following attributes:
- proximity_type
The proximity type used (
"similarity"or"dissimilarity").- method
The silhouette calculation method supplied in
method.- average
Character — the averaging method:
"crisp","fuzzy", or"median".
See Also
Silhouette, softSilhouette, dbSilhouette, cerSilhouette, is.Silhouette, plotSilhouette
Examples
# Create a simple crisp Silhouette object (3 columns)
cluster_assignments <- c(1, 1, 2, 2, 3, 3)
neighbor_clusters <- c(2, 2, 1, 1, 1, 1)
silhouette_widths <- c(0.8, 0.7, 0.6, 0.9, 0.5, 0.4)
sil_obj <- getSilhouette(
cluster = cluster_assignments,
neighbor = neighbor_clusters,
sil_width = silhouette_widths,
proximity_type = "dissimilarity",
method = "medoid",
average = "crisp"
)
sil_obj
# Create a fuzzy Silhouette object with weights (4 columns)
weights <- c(0.9, 0.8, 0.7, 0.95, 0.6, 0.5)
sil_fuzzy <- getSilhouette(
cluster = cluster_assignments,
neighbor = neighbor_clusters,
sil_width = silhouette_widths,
weight = weights,
proximity_type = "similarity",
method = "pac",
average = "fuzzy"
)
sil_fuzzy
# Custom method name
sil_custom <- getSilhouette(
cluster = cluster_assignments,
neighbor = neighbor_clusters,
sil_width = silhouette_widths,
proximity_type = "dissimilarity",
method = "my_custom_method",
average = "crisp"
)
sil_custom
Check if Object is of Class Silhouette
Description
Tests whether an object is of class "Silhouette". This function checks both the class inheritance and the expected structure of a Silhouette object.
Usage
is.Silhouette(x, strict = FALSE)
Arguments
x |
An object to test |
strict |
Logical; if TRUE, performs additional structural validation beyond just class checking (default: FALSE) |
Details
When strict = FALSE, the function only checks if the object inherits
from the "Silhouette" class.
When strict = TRUE, the function additionally validates:
Object is a data frame
Has required columns: cluster, neighbor, sil_width
Has required attributes: proximity_type, method, average
Column types are appropriate (integer for cluster/neighbor, numeric for sil_width)
The Silhouette object attributes are validated as follows:
-
proximity_type: Must be one of "dissimilarity" or "similarity" -
average: Must be one of "crisp", "fuzzy", or "median" -
method: any string, orNA
Value
Logical; TRUE if the object is of class "Silhouette", FALSE otherwise
See Also
Silhouette, softSilhouette, dbSilhouette, cerSilhouette, getSilhouette, plotSilhouette
Examples
# Create a Silhouette object
cluster_assignments <- c(1, 1, 2, 2, 3, 3)
neighbor_clusters <- c(2, 2, 1, 1, 1, 1)
silhouette_widths <- c(0.8, 0.7, 0.6, 0.9, 0.5, 0.4)
sil_obj <- getSilhouette(
cluster = cluster_assignments,
neighbor = neighbor_clusters,
sil_width = silhouette_widths,
proximity_type = "dissimilarity",
method = "medoid",
average = "crisp"
)
# Test if object is Silhouette
is.Silhouette(sil_obj) # TRUE
is.Silhouette(sil_obj, strict = TRUE) # TRUE
# Test with non-Silhouette objects
is.Silhouette(data.frame(a = 1, b = 2)) # FALSE
is.Silhouette(matrix(1:10, ncol = 2)) # FALSE
is.Silhouette(list(a = 1, b = 2)) # FALSE
is.Silhouette(NULL) # FALSE
Plot Silhouette Analysis Results
Description
Creates a silhouette plot for visualizing the silhouette widths of clustering results, with bars colored by cluster and an optional summary of cluster statistics in legend.
Usage
plotSilhouette(
x,
label = FALSE,
summary.legend = TRUE,
grayscale = FALSE,
linetype = c("dashed", "solid", "dotted", "dotdash", "longdash", "twodash"),
...
)
Arguments
x |
An object of class |
label |
Logical; if |
summary.legend |
Logical; if |
grayscale |
Logical; if |
linetype |
Character or numeric value specifying the type of line to be used for the horizontal reference line indicating the average silhouette width. Accepts standard ggplot2 linetype values, such as:
Defaults to |
... |
Additional arguments passed to |
Details
The Silhouette plot displays the silhouette width (sil_width) for each observation, grouped by cluster, with bars sorted by cluster and descending silhouette width. The summary.legend option adds cluster sizes and average silhouette widths to the legend.
This function is the S3 plot method for objects of class "Silhouette", typically produced by the Silhouette, softSilhouette, dbSilhouette or cerSilhouette functions in this package. It also supports objects of the following classes, with silhouette information extracted from their respective component:
-
"eclust": Produced byeclustfrom the factoextra package. -
"hcut": Produced byhcutfrom the factoextra package. -
"pam": Produced bypamfrom the cluster package. -
"clara": Produced byclarafrom the cluster package. -
"fanny": Produced byfannyfrom the cluster package. -
"silhouette": Produced bysilhouettefrom the cluster package orsilhouettefrom the drclust package.
For these classes ("eclust", "hcut", "pam", "clara", "fanny", "silhouette"), users should explicitly call plotSilhouette() (e.g., plotSilhouette(pam_result)) to ensure the correct method is used, as the generic plot() may not dispatch to this function for these objects.
The design of the plot follows fviz_silhouette in the factoextra package (Kassambara and Mundt, 2020), extended to the "Silhouette" class, fuzzy and median averages, and grayscale output.
Value
A ggplot2 object representing the Silhouette plot.
References
Kassambara, A., & Mundt, F. (2020). factoextra: Extract and Visualize the Results of Multivariate Data Analyses. R package version 1.0.7. https://CRAN.R-project.org/package=factoextra
Rousseeuw, P. J. (1987). Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20, 53–65. doi:10.1016/0377-0427(87)90125-7
See Also
Silhouette, softSilhouette, dbSilhouette, cerSilhouette, getSilhouette, is.Silhouette
Examples
data(iris)
# Crisp Silhouette with k-means
out <- kmeans(iris[, -5], 3)
if (requireNamespace("proxy", quietly = TRUE)) {
d_km <- proxy::dist(iris[, -5], out$centers)
plot(Silhouette(d_km))
plot(Silhouette(d_km), palette = "jco")
}
# Fuzzy Silhouette with ppclust::fcm
if (requireNamespace("ppclust", quietly = TRUE)) {
library(ppclust)
out_fuzzy <- Silhouette(
prox_matrix = "d",
proximity_type = "dissimilarity",
prob_matrix = "u",
clust_fun = ppclust::fcm,
x = iris[, 1:4],
centers = 3,
sort = TRUE
)
plot(out_fuzzy, summary.legend = FALSE, grayscale = TRUE)
} else {
message("Install 'ppclust': install.packages('ppclust')")
}
# Silhouette plot for pam clustering
if (requireNamespace("cluster", quietly = TRUE)) {
library(cluster)
pam_result <- pam(iris[, 1:4], k = 3)
plotSilhouette(pam_result)
}
# Silhouette plot for clara clustering
if (requireNamespace("cluster", quietly = TRUE)) {
clara_result <- clara(iris[, 1:4], k = 3)
plotSilhouette(clara_result)
}
# Silhouette plot for fanny clustering
if (requireNamespace("cluster", quietly = TRUE)) {
fanny_result <- fanny(iris[, 1:4], k = 3)
plotSilhouette(fanny_result)
}
# Example using base silhouette() object
if (requireNamespace("cluster", quietly = TRUE)) {
sil <- silhouette(pam_result)
plotSilhouette(sil)
}
# Silhouette plot for eclust clustering
if (requireNamespace("factoextra", quietly = TRUE)) {
library(factoextra)
eclust_result <- eclust(iris[, 1:4], "kmeans", k = 3, graph = FALSE)
plotSilhouette(eclust_result)
}
# Silhouette plot for hcut clustering
if (requireNamespace("factoextra", quietly = TRUE)) {
hcut_result <- hcut(iris[, 1:4], k = 3)
plotSilhouette(hcut_result)
}
# Silhouette plot for a drclust partition (3 clusters, 2 components)
if (requireNamespace("drclust", quietly = TRUE)) {
library(drclust)
iris_mat <- as.matrix(iris[,-5])
drclust_out <- dpcakm(iris_mat, 3, 2)
d <- silhouette(iris_mat, drclust_out)
plotSilhouette(d$cl.silhouette)
}
Calculate Silhouette Width for Soft Clustering Algorithms
Description
Computes silhouette widths for soft clustering results by interpreting cluster membership probabilities (or their transformations) as proximity measures (Bhat Kapu and Kiruthika, 2024). Although originally designed for evaluating clustering quality within a method, this adaptation allows heuristic comparison across soft clustering algorithms using average silhouette widths.
Usage
softSilhouette(
prob_matrix,
prob_type = c("pp", "nlpp", "pd"),
method = c("pac", "medoid"),
average = c("crisp", "fuzzy", "median"),
a = 2,
sort = FALSE,
print.summary = FALSE,
clust_fun = NULL,
...,
cluster = NULL,
clust_args = list()
)
Arguments
prob_matrix |
A numeric matrix where rows represent observations and columns represent cluster membership probabilities. Rows must sum to 1. If |
prob_type |
Character string specifying the type transformation of membership matrix considered as proximity matrix in
Defaults to |
method |
Character string specifying the normalization of the silhouette width. Options are |
average |
Character string specifying the method for computing the average silhouette width. Options are:
Defaults to |
a |
Non-negative number, the exponent |
sort |
Logical; if |
print.summary |
Logical; if |
clust_fun |
Optional S3 or S4 function object or function as character string specifying a clustering function that produces the proximity measure matrix. For example, |
... |
Additional arguments passed to |
cluster |
Optional integer vector of length |
clust_args |
A list of further arguments for |
Details
With the names used by Bhat Kapu and Kiruthika (2024): prob_type = "pp" with
method = "pac" is the PACS (the silhouette based on the probability of the
alternative class of Raymaekers and Rousseeuw, 2022, applied to posterior probabilities), "pp" with "medoid" the PPS (posterior probability silhouette),
"nlpp" with "medoid" the NLPPS (negative log of posterior probability
silhouette), and "pd" with "medoid" the PDS (probability density
silhouette). The certainty silhouette (CeS) and the density-based silhouette (DBS) are
computed by cerSilhouette and dbSilhouette.
The membership matrix is transformed according to prob_type and the widths are
computed as in Silhouette. The PPS and PACS treat the posterior
probabilities as a similarity proximity measure, the NLPPS turns them into a dissimilarity
through the negative logarithm, and the PDS treats the component densities as a similarity
(Bhat Kapu and Kiruthika, 2024). With method = "pac", the "pp" width is
(\gamma_{ic_i} - \gamma_{i\nu_i})/(\gamma_{ic_i} + \gamma_{i\nu_i}); with
"medoid" it is 1 - \gamma_{i\nu_i}/\gamma_{ic_i}.
For a finite mixture f(x) = \sum_k \pi_k f_k(x) fitted by EM,
\gamma_{ik} = \pi_k f_k(x_i)/f(x_i) and, at convergence, \hat\pi_k is the
fitted mixing proportion, so \gamma_{ik}/\hat\pi_k = f_k(x_i)/f(x_i) (up to the
convergence tolerance). The PDS width of Equation (22) of Bhat Kapu and Kiruthika (2024),
(f_{c_i}(x_i) - \max_{k \neq c_i} f_k(x_i)) / \max(f_{c_i}(x_i), \max_{k \neq c_i} f_k(x_i)),
is therefore computed from the memberships alone, and can be used with any soft clustering
algorithm.
A zero probability gives an infinite "nlpp" dissimilarity. The width is then 1 if
only the neighboring membership is zero, -1 if only the assigned membership is zero, and
0 if both are zero.
Value
A data frame of class "Silhouette" with one row per observation and columns
cluster, neighbor, weight (only when average = "fuzzy") and
sil_width. The object includes the following attributes:
- proximity_type
The proximity type used (
"similarity"or"dissimilarity").- method
The silhouette calculation method used (
"medoid"or"pac").- average
Character — the averaging method:
"crisp","fuzzy", or"median".
Choosing the number of clusters
The average of these widths over k = 2, 3, \ldots typically starts near 1 and
decreases. Bhat Kapu and Kiruthika (2024) read the number of clusters from the reverse
elbow drop of this curve, the value of k after which the average drops sharply,
rather than from its maximum. Which k is suggested depends on the algorithm that
produced the memberships, on the width and on the averaging, so report all three with the
result.
Cluster labels
The labels are cluster when supplied. Otherwise, for "pp" and "nlpp",
each observation is assigned to the cluster with its largest membership, the partition a
soft clustering algorithm reports, and the widths are non-negative. For "pd", each
observation is assigned to the cluster with its largest density
\gamma_{ik}/\hat\pi_k, as in the definition of the PDS, so the PDS widths lie in
[0, 1] for every average. The largest density and the largest membership coincide
when the clusters have equal estimated weights; otherwise the PDS can place an observation
in a smaller cluster than its largest membership does. The neighboring cluster is the
closest other cluster in the transformed matrix.
Ties are resolved in favor of the first column; see Silhouette.
Averages
If average = "crisp", the crisp silhouette index (CS) is the mean width:
CS = \frac{1}{n} \sum_{i=1}^{n} S(x_i)
summarizing overall clustering quality.
If average = "fuzzy" and prob_matrix is provided, denoted as \Gamma = [\gamma_{ik}]_{n \times K},
with \gamma_{ik} representing the probability of observation i belonging to cluster k,
the fuzzy silhouette index (FS) of Campello and Hruschka (2006) is:
FS = \frac{\sum_{i=1}^{n} w_i S(x_i) }{\sum_{i=1}^{n} w_i},
\qquad w_i = ( \gamma_{i(1)} - \gamma_{i(2)} )^{\alpha},
where \gamma_{i(1)} and \gamma_{i(2)} are the largest and second largest
memberships of observation i, and \alpha (the a argument) controls the
emphasis on confident assignments (\alpha \ge 0; \alpha = 0 gives the crisp
average). The weight w_i is stored in the weight column and does not depend on
the cluster labels; it equals Equation (8) of Bhat Kapu and Kiruthika (2024) when each
observation is assigned to its largest membership. If all weights are zero, the fuzzy
average is undefined (NaN) and a warning is given.
If average = "median", the median width is reported.
Missing widths are removed before averaging.
References
Raymaekers, J., & Rousseeuw, P. J. (2022). Silhouettes and quasi residual plots for neural nets and tree-based classifiers. Journal of Computational and Graphical Statistics, 31(4), 1332–1343. doi:10.1080/10618600.2022.2050249
Campello, R. J., & Hruschka, E. R. (2006). A fuzzy extension of the silhouette width criterion for cluster analysis. Fuzzy Sets and Systems, 157(21), 2858–2875. doi:10.1016/j.fss.2006.07.006
Bhat Kapu, S., & Kiruthika. (2024). Some density-based silhouette diagnostics for soft clustering algorithms. Communications in Statistics: Case Studies, Data Analysis and Applications, 10(3-4), 221-238. doi:10.1080/23737484.2024.2408534
See Also
Silhouette, dbSilhouette, cerSilhouette, calSilhouette, getSilhouette, is.Silhouette, plotSilhouette
Examples
# A small membership matrix
prob <- matrix(c(0.90, 0.10,
0.60, 0.40,
0.20, 0.80,
0.45, 0.55), ncol = 2, byrow = TRUE)
softSilhouette(prob)
softSilhouette(prob, prob_type = "nlpp", average = "fuzzy")
# Compare two soft clustering algorithms on the iris data
data(iris)
if (requireNamespace("ppclust", quietly = TRUE)) {
fcm_result <- ppclust::fcm(iris[, 1:4], 3)
out_fcm <- softSilhouette(prob_matrix = fcm_result$u, print.summary = TRUE)
plot(out_fcm)
fcm2_result <- ppclust::fcm2(iris[, 1:4], 3)
out_fcm2 <- softSilhouette(prob_matrix = fcm2_result$u)
c(fcm = summary(out_fcm, print.summary = FALSE)$avg.width,
fcm2 = summary(out_fcm2, print.summary = FALSE)$avg.width)
} else {
message("Install 'ppclust' to run this example: install.packages('ppclust')")
}