Clustergram {shipunov}

Clustergram: visualize the cluster structure

Description

Plot which shows cluster memberships and distances when clusters numbers increases

Usage

  Clustergram(data, maxcl=ncol(data)*2, nboot=FALSE, method="kmeans",
  m.dist="euclidean", m.hclust="complete", plot=TRUE, broom=4e-3,
  col="gray", ...)

Arguments

data

Data, typically data frame

maxcl

Maximal number of clusters, default is number of columns times 2; minimal number of clusters is 2

nboot

Either 'FALSE' (no bootstrap, default) or number of bootstrap runs

method

Either 'kmeans' or 'hclust'

m.dist

If method='hclust', method to calculate distances, see ?dist

m.hclust

If method='hclust', method to clusterize, see ?hclust

plot

Plot?

broom

Extent to which spread lines, default is 4e-3

col

Color of lines

...

Further arguments to plot()

Details

Clustergram shows how cluster members are assigned to clusters as the number of clusters increases. This graph is useful in exploratory analysis for non-hierarchical clustering algorithms like k-means and for hierarchical cluster algorithms when the number of observations is large enough to make dendrograms impractical (from Schonlau, 2004; see also www.schonlau.net).

One application is to use clustergram to determine the optimal number of clusters. Basic idea is that you look for the point (number of clusters) where more clusters do not significanly change the picture (i.e., do not add more information) The best number of clusters is _near_ that point (see examples).

See also Martin Fleischmann (martinfleischmann.net) for practical explanation and scikit-learn 'clustergram' Python package.

Clustergram() code based on simplified and optimized Tal Galili's github 'clustergram' code.

Author(s)

Alexey Shipunov

References

Schonlau M. 2004. Visualizing non-hierarchical and hierarchical cluster analyses with clustergrams. Computational Statistics 19, 95-111.

See Also

dist, hclust, kmeans

Examples

set.seed(250)
aa <- matrix(rnorm(20000), nrow=100)
## maximal number of clusters is less than default
## line color is like in scikit-learn
## larger "broom" so lines are a bit broader
Clustergram(aa, maxcl=5, col="#3B6E8C", broom=2e-2, main="Artificial data, homogeneous")
aa[1:60, ] <- aa[1:60, ] + 0.7
aa[1:20, ] <- aa[1:20, ] + 0.4
Clustergram(aa, maxcl=5, col="#F29528", broom=2e-2,
 main="Artificial data, heterogeneous, 3 clusters")

## Clustergram() invisibly outputs matrix of clusters
ii <- Clustergram(iris[, -5], main=expression(bolditalic("Iris")*bold(" data")))
head(ii)
## Hierarchical clustering instead of kmeans
Clustergram(iris[, -5], method="hclust", m.hclust="average", main="Iris, UPGMA")
## Bootstrap. Resulted PDF could be opening slowly, use raster images in that case
Clustergram(iris[, -5], nboot=100, col=adjustcolor("darkgray", alpha.f=0.3),
 main="Iris, kmeans, nboot 100")