wskm (version 1.4.40)

fgkm: Feature Group Weighting K-Means for Subspace clustering

Description

Perform an feature group weighting subspace k-means.

Usage

fgkm(x, centers, groups, lambda, eta, maxiter=100, delta=0.000001,
       maxrestart=10,seed=-1)

Arguments

x

numeric matrix of observations and features.

centers

target number of clusters or the initial centers for clustering.

groups

a string give the group information, formatted as "0,1,2,4;3,5;6,7,8" or "0-2,4;3,5;6-8", where ";" defines a group; or a vector of length of features, each element of the vector indicates the group of the feature. For example, c(1,1,1,2,1,2,3,3,3) is the same as "0-2,4;3,5;6-8", or even c("a","a","a","b","a","b","c","c","c").

lambda

parameter of feature weight distribution.

eta

parameter of group weight distribution.

delta

maximum change allowed between iterations for convergence.

maxiter

maximum number of iterations.

maxrestart

maximum number of restarts. Default is 10 so that we stand a good chance of getting a full set of clusters. Normally, any empty clusters that result are removed from the result, and so we may obtain fewer than k clusters if we don't allow restarts(i.e., maxrestart=0). If < 0, then there is no limit on the number of restarts and we are much likely to get a full set of k clusters.

seed

random seed. If it was set below 0, then a randomly generated number will be assigned.

Value

Return an object of class "kmeans" and "fgkm", compatible with other function that work with kmeans objects, such as the 'print' method. The object is a list with the following components in addition to the components of the kmeans object:

cluster

A vector of integer (from 1:k) indicating the cluster to which each point is allocated.

centers

A matrix of cluster centers.

featureWeight

A matrix of weights recording the relative importance of each feature for each cluster.

groupWeight

A matrix of group weights recording the relative importance of each feature group for each cluster.

iterations

This report on the number of iterations before termination. Check this to see whether the maxiters was reached. If so then teh algorithm may not be converging, and thus the resulting clustering may not be particularly good.

restarts

The number of times the clustering restarted because of a disappearing cluster resulting from one or more k-means having no observations associated with it. An number here greater than zero indicates that the algorithm is not converging on a clustering for the given k. It is recommended that k be reduced.

totalIterations

The total number of iterations over all restarts.

totolCost

The total cost calculated in the cost function.

Details

The feature group weighting k-means clustering algorithm is a extension to ewkm, which itself is a soft subspace clustering method.

The algorithm weights subspaces in both feature groups and individual features.

Always check the number of iterations, the number of restarts, and the total number of iterations as they give a good indication of whether the algorithm converged.

As with any distance based algorithm, be sure to rescale your numeric data so that large values do not bias the clustering. A quick rescaling method to use is scale.

References

Xiaojun Chen, Yunming Ye, Xiaofei Xu and Joshua Zhexue Huang (2012). A Feature Group Weighting Method for Subspace Clustering of High-Dimensional Data. Pattern Recognition, 45(1), 434--446.

See Also

kmeans ewkm twkm

Examples

# NOT RUN {
# The data fgkm.sample has 600 objects and 50 dimensions.
# Scale the data before clustering
x <- scale(fgkm.sample)

# Group information is formated as below.
# Each group is separated by ';'.
strGroup <- "0-9;10-19;20-49"
groups <- c(rep(0, 10), rep(1, 10), rep(2, 30))

# Use the fgkm algorithm.
myfgkm <- fgkm(x, 3, strGroup, 3, 1, seed=19)
myfgkm2 <- fgkm(x, 3, groups, 3, 1, seed=19)
all.equal(myfgkm, myfgkm2)

# You can print the clustering result now.
myfgkm$cluster
myfgkm$featureWeight
myfgkm$groupWeight
myfgkm$iterations
myfgkm$restarts
myfgkm$totiters
myfgkm$totss

# Use a cluster validation method from package 'fpc'.

# real.cluster is the real class label of the data 'fgkm.sample'.
real.cluster <- rep(1:3, each=200)

# cluster.stats() computes several distance based statistics.
kmstats <- cluster.stats(d=dist(x), as.integer(myfgkm$cluster), real.cluster)

# corrected Rand index
kmstats$corrected.rand

# variation of information (VI) index
kmstats$vi

# }