kmeans function. It creates
several partitions forming a cascade from a small to a large number of
groups.cascadeKM(data, inf.gr, sup.gr, iter = 100, criterion = "calinski")cIndexKM(y, x, index = "all")
## S3 method for class 'cascadeKM':
plot(x, min.g, max.g, grpmts.plot = TRUE,
sortg = FALSE, gridcol = NA, ...)
"calinski", which refers to
the Calinski-Harabasz (1974) criterion. The simple structure index
("ssi") is also available. Other indice"kmeans" returned by a clustering algorithm
such as kmeansplot"calinski" and "ssi".
Type "all" to obtain both indices.
Abbreviations of these names are also accepted.TRUE or FALSE).x, although not in the graph. If NA,
which is the default value, removes the grid lines.cascadeKM returns an object of class
cascadeKM with items:inf.gr to $K$ =
sup.gr.cIndex returns a vector with the index values. The
maximum value of these indices is supposed to indicate the best
partition. These indices work best with groups of equal sizes. When
the groups are not of equal sizes, one should not put too much faith
in the maximum of these indices, and also explore the groups
corresponding to other values of $K$.kmeans. Most
of the work is performed by function cIndex which is based on the
clustIndex function (package "calinski", which refers to the well-known
Calinski-Harabasz (1974) criterion. The other available index is the
simple structure index "ssi" (Dolnicar et al. 1999).
In the case of groups of equal
sizes, "calinski" is generally a good criterion to indicate the
correct number of groups. Users should not take its indications
literally when the groups are not equal in size. Type "all" to
obtain both indices. The indices are defined as:
[object Object],[object Object] In a simulation study, Milligan and Cooper (1985) found
that the Calinski-Harabasz criterion recovered the correct number of
groups the most often. We recommend this criterion because, if the
groups are of equal sizes, the maximum value of "calinski"
usually indicates the correct number of groups. Another available
index is the simple structure index "ssi". Users should not
take the indications of these indices literally when the groups are
not equal in size and explore the groups corresponding to other values
of $K$.
Function cascadeKM has a plot method. Two plots are
produced. The graph on the left has the objects in
abscissa and the number of groups in ordinate. The groups are
represented by colours. The graph on the right shows the values of the
criterion ("calinski" or "ssi") for determining the best
partition. The highest value of the criterion is marked in red. Points
marked in orange, if any, indicate partitions producing an increase in
the criterion value as the number of groups increases; they may
represent other interesting partitions.
If sortg=TRUE, the objects are reordered by the following
procedure: (1) a simple matching distance matrix is computed among the
objects, based on the table of K-means assignments to groups, from
$K$ = min.g to $K$ = max.g. (2) A principal
coordinate analysis (PCoA, Gower 1966) is computed on the centred
distance matrix. (3) The first principal coordinate is used as the new
order of the objects in the graph. A simplified algorithm is used to
compute the first principal coordinate only, using the iterative
algorithm described in Legendre & Legendre (2012). The
full distance matrix among objects is never computed; this avoids
the problem of storing it when the number of objects is
large. Distance values are computed as they are needed by the
algorithm.
Dolnicar, S., K. Grabler and J. A. Mazanec. 1999. A tale of three cities: perceptual charting for analyzing destination images. Pp. 39-62 in: Woodside, A. et al. [eds.] Consumer psychology of tourism, hospitality and leisure. CAB International, New York.
Gower, J. C. 1966. Some distance properties of latent root and vector methods used in multivariate analysis. Biometrika 53: 325-338. Legendre, P. & L. Legendre. 2012. Numerical ecology, 3rd English edition. Elsevier Science BV, Amsterdam. Milligan, G. W. & M. C. Cooper. 1985. An examination of procedures for determining the number of clusters in a data set. Psychometrika 50: 159-179.
Weingessel, A., Dimitriadou, A. and Dolnicar, S. An Examination
Of Indexes For Determining The Number Of Clusters In Binary Data
Sets,
kmeans, clustIndex.# Partitioning a (10 x 10) data matrix of random numbers
mat <- matrix(runif(100),10,10)
res <- cascadeKM(mat, 2, 5, iter = 25, criterion = 'calinski')
toto <- plot(res)
# Partitioning an autocorrelated time series
vec <- sort(matrix(runif(30),30,1))
res <- cascadeKM(vec, 2, 5, iter = 25, criterion = 'calinski')
toto <- plot(res)
# Partitioning a large autocorrelated time series
# Note that we remove the grid lines
vec <- sort(matrix(runif(1000),1000,1))
res <- cascadeKM(vec, 2, 7, iter = 10, criterion = 'calinski')
toto <- plot(res, gridcol=NA)Run the code above in your browser using DataLab