qVarSel (version 1.1)

qVarSel-package: Selecting Variables for Clustering and Classification

Description

For a given data matrix A and cluster centers/prototypes collected in the matrix P, the functions described here select a subset of variables Q that mostly explains/justifies P as prototypes. The functions are useful to reduce the dimension of the data for classification as they discard masking variables for clustering.

Author

Stefano Benati

Maintainer: Stefano Benati <[e-mail]>

Details

Package:	qVarSel
Type:	Package
Version:	1.1
Date:	2024-11-15
License:	gpl-2

The package is useful to reduce the variable dimension for clustering. The example below shows the sequence of the operations. First, k-means can applied to the whole data sets, to calculate prototypes P. Then, distances between units U and P are calculated ans stored in a matrix D. Then, apply package subroutine q-VarSelH to select the most important variables. Apply EM optimization on data D for full clustering parameters estimation.

References

S. Benati, S. Garcia Quiles, J. Puerto "Mixed Integer Linear Programming and heuristic methods for feature selection in clustering", Journal of the Operational Research Society, 69:9, (2018), pp. 1379-1395

Examples

# Simulated data with 100 units, 10 true variables,
# 10 masking variables, 2 hidden clusters

n1 = 50
n2 = 50
n_true_var = 10
n_mask_var = 10
g1 = matrix(rnorm(n1*n_true_var, 0, 1), ncol = n_true_var)
g2 = matrix(rnorm(n2*n_true_var, 2, 1), ncol = n_true_var)
m1 = matrix(runif((n1 + n2)*n_mask_var, min = 0, max = 5), ncol = n_mask_var)
a = cbind(rbind(g1, g2), m1)

## calculate data prototypes using k-means

sl2 <- kmeans(a, 2, iter.max = 100, nstart = 2)
p = sl2$centers

## calculate distances between observations and prototypes
## Remark: d is a 3-dimensions matrix

d = PrtDist(a, p)

## Select 10 most representative variables, use heuristic

lsH <- qVarSelH(d, 10, maxit = 200)

# reduce the dimension of a

sq = 1:(dim(a)[2])
vrb = sq[lsH$x > 0.01]
a_reduced = a[ ,vrb]

# use the EM methodology for effcient clustering on the reduced data

require(mclust)
sl1 <- Mclust(a_reduced, G = 2, modelName = "VVV")