bclust: Bagged Clustering

Description

Cluster the data in x using the bagged clustering algorithm. A partitioning cluster algorithm such as kmeans is run repeatedly on bootstrap samples from the original data. The resulting cluster centers are then combined using the hierarchical cluster algorithm hclust.

Usage

bclust(x, centers=2, iter.base=10, minsize=0,
       dist.method="euclidian",
       hclust.method="average", base.method="kmeans",
       base.centers=20, verbose=TRUE,
       final.kmeans=FALSE, docmdscale=FALSE,
       resample=TRUE, weights=NULL, maxcluster=base.centers, ...)
hclust.bclust(object, x, centers, dist.method=object$dist.method,
              hclust.method=object$hclust.method, final.kmeans=FALSE,
              docmdscale = FALSE, maxcluster=object$maxcluster)
## S3 method for class 'bclust':
plot(x, maxcluster=x$maxcluster, main, ...)
centers.bclust(object, k)
clusters.bclust(object, k, x=NULL)

Arguments

Matrix of inputs (or object of class "bclust" for plot).

centers, k

Number of clusters.

iter.base

Number of runs of the base cluster algorithm.

minsize

Minimum number of points in a base cluster.

dist.method

Distance method used for the hierarchical clustering, see dist for available distances.

hclust.method

Linkage method used for the hierarchical clustering, see hclust for available methods.

base.method

Partitioning cluster method used as base algorithm.

base.centers

Number of centers used in each repetition of the base method.

verbose

Output status messages.

final.kmeans

If TRUE, a final kmeans step is performed using the output of the bagged clustering as initialization.

docmdscale

Logical, if TRUE a cmdscale result is included in the return value.

resample

Logical, if TRUE the base method is run on bootstrap samples of x, else directly on x.

weights

Vector of length nrow(x), weights for the resampling. By default all observations have equal weight.

maxcluster

Maximum number of clusters memberships are to be computed for.

object

Object of class "bclust".

main

Main title of the plot.

...

Optional arguments top be passed to the base method in bclust, ignored in plot.

Value

bclust and hclust.bclust return objects of class "bclust" including the components
hclustReturn value of the hierarchical clustering of the collection of base centers (Object of class "hclust").
clusterVector with indices of the clusters the inputs are assigned to.
centersMatrix of centers of the final clusters. Only useful, if the hierarchical clustering method produces convex clusters.
allcentersMatrix of all iter.base * base.centers centers found in the base runs.

Details

First, iter.base bootstrap samples of the original data in x are created by drawing with replacement. The base cluster method is run on each of these samples with base.centers centers. The base.method must be the name of a partitioning cluster function returning a list with the same components as the return value of kmeans.

This results in a collection of iter.base * base.centers centers, which are subsequently clustered using the hierarchical method hclust. Base centers with less than minsize points in there respective partitions are removed before the hierarchical clustering.

The resulting dendrogram is then cut to produce centers clusters. Hence, the name of the argument centers is a little bit misleading as the resulting clusters need not be convex, e.g., when single linkage is used. The name was chosen for compatibility with standard partitioning cluster methods such as kmeans.

A new hierarchical clustering (e.g., using another hclust.method) re-using previous base runs can be performed by running hclust.bclust on the return value of bclust.

References

Friedrich Leisch. Bagged clustering. Working Paper 51, SFB ``Adaptive Information Systems and Modeling in Economics and Management Science'', August 1999. http://www.ci.tuwien.ac.at/~leisch

Examples

Run this code

data(iris)
bc1 <- bclust(iris[,1:4], 3, base.centers=5)
plot(bc1)

table(clusters.bclust(bc1, 3))
centers.bclust(bc1, 3)

Run the code above in your browser using DataCamp Workspace