Skip to contents

cluster_choice() clusters the same sequences under every combination of the cluster counts, dissimilarities and algorithms given, each with build_clusters(), and returns one row per combination with its average silhouette width, mean within-cluster distance and cluster-size balance. It is the distance-based counterpart of compare_mmm(), which compares mixtures of Markov chains across k. "all" expands dissimilarity or method to every supported option.

The silhouette of a sequence (Rousseeuw, 1987) compares its mean distance to its own cluster with that to the nearest other cluster: near 1 it sits well inside its cluster, near 0 between two, and below 0 it is closer to another cluster. Rows that share a dissimilarity compare k and algorithms on one distance scale; across dissimilarities the silhouettes and within-cluster distances rest on different distances and compare less directly. The row with the highest silhouette is marked, but that is not a decision: also weigh the cluster sizes (min_size, size_ratio), as a high silhouette can come with a cluster of very few sequences; then fit the chosen configuration with build_clusters() and read it with its summary(), plot() and cluster_diagnostics().

Usage

cluster_choice(
  data,
  k = 2:5,
  dissimilarity = "hamming",
  method = "ward.D2",
  ...
)

# S3 method for class 'cluster_choice'
print(x, digits = 3L, ...)

# S3 method for class 'cluster_choice'
summary(object, ...)

# S3 method for class 'cluster_choice'
plot(
  x,
  type = c("auto", "lines", "bars", "heatmap", "tradeoff", "facet"),
  abbrev = FALSE,
  combined = TRUE,
  ...
)

Arguments

data

The sequences, passed to build_clusters() unchanged: any input it takes, such as a wide data frame of sequences or a network built from sequences.

k

The cluster counts to try. Default 2:5. Each must be a whole finite number, or an error is raised; build_clusters() also requires each to be at least 2 and at most the number of sequences minus 1.

dissimilarity

The dissimilarity measures to try. Default "hamming". "all" expands to every supported measure: "hamming", "osa", "lv", "dl", "lcs", "qgram", "cosine", "jaccard" and "jw", and may stand beside named ones (each is fitted once).

method

The clustering algorithms to try. Default "ward.D2" (where build_clusters() defaults to "pam"). "all" expands to every supported algorithm: "pam", "ward.D2", "ward.D", "complete", "average", "single", "mcquitty", "median" and "centroid".

...

Further arguments of build_clusters(), used for every combination: weighted, lambda, q, p, seed, na_syms, covariates and estimator; any other name raises an error. weighted = TRUE works only with dissimilarity = "hamming" and is refused before any fit when the sweep holds another measure. The print(), summary() and plot() methods take no further arguments and raise an error on any.

x

For the print() and plot() methods: a cluster_choice.

digits

For the print() method: the decimal places of the silhouette, within-cluster distance and size ratio. Default 3. Must be a single non-negative whole number.

object

For the summary() method: a cluster_choice.

type

For the plot() method: the kind of plot, one of "auto" (default), "lines", "bars", "heatmap", "tradeoff" and "facet" (see Plot types). A type the sweep cannot support raises an error that names the alternatives, and so does a table with a single configuration, which has nothing to compare.

abbrev

For the plot() method: whether to shorten the dissimilarity and method names (e.g. "hamming" to "ham", "ward.D2" to "wD2") on the axes, point labels and legend. Default FALSE.

combined

For plot(type = "facet"): TRUE (default) draws one panel per method in a single plot; FALSE returns a named list of one-method plots instead. Ignored by the other types, but must still be a single logical value.

Value

A cluster_choice (a data frame subclass) with one row per combination, k varying fastest, then dissimilarity, then method. print() shows the table, with only the settings that vary as columns, and flags the row with the highest silhouette in a best column; summary() returns the full table as a plain data frame with a character best column that holds "silhouette" on that row; plot() draws the silhouette across the swept settings (see Plot types) and returns the ggplot invisibly (with type = "facet" and combined = FALSE, a named list of ggplots, returned invisibly and not drawn). The print() method returns its input invisibly. The columns:

k, dissimilarity, method

The configuration of the row.

silhouette

The average silhouette width over all sequences, computed by build_clusters() (from cluster::pam() for "pam", with cluster::silhouette() otherwise).

mean_within_dist

The mean distance between members of the same cluster, averaged over clusters weighted by their size (a one-member cluster counts as 0), in the units of the row's dissimilarity.

min_size, max_size, size_ratio

The sizes of the smallest and the largest cluster, and their ratio max_size / min_size.

Plot types

The plots show the values only: they carry no "best" marker and no recommendation. plot() needs at least one swept setting.

"auto"

The default. Picks a type from the swept settings: "lines" when k is swept with at most one other setting, "bars" when only dissimilarity or only method is swept, "heatmap" when both are swept at a single k, and "facet" when all three are swept.

"lines"

The silhouette across k, one line per value of the other swept setting. When only k is swept, mean_within_dist is drawn as a second line on the same axis. With all three swept, the lines follow dissimilarity and the methods share them; use "facet" instead. Needs k swept.

"bars"

Horizontal bars of the silhouette per value of the swept setting, or per combination when several are swept, sorted by silhouette, with the values printed.

"heatmap"

Tiles of the silhouette over dissimilarity and method, with the values printed. Needs both swept, and is meant for a single k.

"tradeoff"

The silhouette against size_ratio, each point labelled with its swept settings.

"facet"

The silhouette across k, one line per dissimilarity and one panel per method. Needs all three swept.

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

See also

build_clusters, compare_mmm for the model-based equivalent, cluster_diagnostics for the post-fit diagnostic surface on a single clustering.

Examples

seqs <- data.frame(V1 = sample(c("A","B","C"), 40, TRUE),
                   V2 = sample(c("A","B","C"), 40, TRUE))
cluster_choice(seqs, k = 2:4)
#> Cluster Choice (sweep: k)
#> 
#>  k silhouette within_dist sizes    ratio best    
#>  2 0.403      0.995       [17, 23] 1.353         
#>  3 0.563      0.711       [11, 17] 1.545 <-- best
#>  4 0.541      0.553       [7, 12]  1.714         
# \donttest{
# Sweep dissimilarities at fixed k
cluster_choice(seqs, k = 3, dissimilarity = c("hamming", "lcs", "jaccard"))
#> Cluster Choice (sweep: dissimilarity)
#> 
#>  dissimilarity silhouette within_dist sizes    ratio best    
#>  hamming       0.563      0.711       [11, 17] 1.545 <-- best
#>  lcs           0.403      1.363       [9, 19]  2.111         
#>  jaccard       0.381      0.619       [5, 28]  5.600         

# Full grid of k x dissimilarity
cluster_choice(seqs, k = 2:4, dissimilarity = c("hamming", "lcs"))
#> Cluster Choice (sweep: k x dissimilarity)
#> 
#>  k dissimilarity silhouette within_dist sizes    ratio best    
#>  2 hamming       0.403      0.995       [17, 23] 1.353         
#>  3 hamming       0.563      0.711       [11, 17] 1.545 <-- best
#>  4 hamming       0.541      0.553       [7, 12]  1.714         
#>  2 lcs           0.368      1.709       [19, 21] 1.105         
#>  3 lcs           0.403      1.363       [9, 19]  2.111         
#>  4 lcs           0.519      1.044       [5, 19]  3.800         

# "all" sentinel
cluster_choice(seqs, k = 3, dissimilarity = "all")
#> Cluster Choice (sweep: dissimilarity)
#> 
#>  dissimilarity silhouette within_dist sizes    ratio best    
#>  hamming       0.563      0.711       [11, 17] 1.545         
#>  osa           0.463      0.714       [10, 17] 1.700         
#>  lv            0.563      0.711       [11, 17] 1.545         
#>  dl            0.463      0.714       [10, 17] 1.700         
#>  lcs           0.403      1.363       [9, 19]  2.111         
#>  qgram         0.381      1.237       [5, 28]  5.600         
#>  cosine        0.381      0.619       [5, 28]  5.600         
#>  jaccard       0.381      0.619       [5, 28]  5.600         
#>  jw            0.718      0.209       [12, 14] 1.167 <-- best
# }