diana2means: 2-Means with Hierarchical Initialization

Description

Split a set of data points into two coherent groups using the k-means algorithm. Instead of random initialization, divisive hierarchical clustering is used to determine initial groups and the corresponding centroids.

Usage

diana2means(mydata, mingroupsize = 5,  ngenes = 50, ignore.genes = 5,  return.cut = FALSE)

Arguments

mydata

either an expression set as defined by the package Biobase or a matrix of expression levels (rows=genes, columns=samples).

mingroupsize

report only splits where both groups are larger than this size.

ngenes

number of genes used to compute cluster quality DLD-score.

ignore.genes

number of best scoring genes to be ignored when computing DLD-scores.

return.cut

logical, whether to return the attributions of samples to groups.

Value

cut: one number out of 0 and 1 per column in the original data, specifying the split attribution.
score: the DLD-score achieved by the split.

Details

This function uses divisive hierarchical clustering (diana) to generate a first split of the data. Thereby, each column of the data matrix is considered to represent a data element. From the thus generated temptative groups, centroids are deduced and used to initialize the k-means clustering algorithm.

For the split optimized by k-means the DLD-score is determined using the ngenes and ignore.genes arguments.

Examples

Run this code

# get golub data
library(vsn)
library(golubEsets)
data(Golub_Merge)

# use 10% most variable genes
e <- exprs(Golub_Merge)
vars <- apply(e, 1, var)
e <- e[vars > quantile(vars,0.9),]

# use diana2means to get splits and scores
diana2means(e)
diana2means(e, return.cut=TRUE)