tclust (version 2.0-3)

ctlcurves: Classification Trimmed Likelihood Curves

Description

The function applies tclust several times on a given dataset while parameters alpha and k are altered. The resulting object gives an idea of the optimal trimming level and number of clusters considering a particular dataset.

Usage

ctlcurves(
  x,
  k = 1:4,
  alpha = seq(0, 0.2, len = 6),
  restr.fact = 50,
  parallel = FALSE,
  trace = 1,
  ...
)

Value

The function returns an S3 object of type ctlcurves containing the following components:

• par A list containing all the parameters passed to this function

• obj An array containing the objective functions values of each computed cluster-solution

• min.weights An array containing the minimum cluster weight of each computed cluster-solution

Arguments

x

A matrix or data frame of dimension n x p, containing the observations (row-wise).

k

A vector of cluster numbers to be checked. By default cluster numbers from 1 to 5 are examined.

alpha

A vector containing the alpha levels to be checked. By default alpha levels from 0 to 0.2 (continuously increased by 0.01), are checked.

restr.fact

The restriction factor passed to tclust.

parallel

A logical value, to be passed further to tclust().

trace

Defines the tracing level, which is set to 1 by default. Tracing level 2 gives additional information on the current iteration.

...

Further arguments (as e.g. restr), passed to tclust

Details

These curves show the values of the trimmed classification (log-)likelihoods when altering the trimming proportion alpha and the number of clusters k. The careful examination of these curves provides valuable information for choosing these parameters in a clustering problem. For instance, an appropriate k to be chosen is one that we do not observe a clear increase in the trimmed classification likelihood curve for k with respect to the k+1 curve for almost all the range of alpha values. Moreover, an appropriate choice of parameter alpha may be derived by determining where an initial fast increase of the trimmed classification likelihood curve stops for the final chosen k. A more detailed explanation can be found in García-Escudero et al. (2011).

References

García-Escudero, L.A.; Gordaliza, A.; Matrán, C. and Mayo-Iscar, A. (2011), "Exploring the number of groups in robust model-based clustering." Statistics and Computing, 21 pp. 585-599, <doi:10.1007/s11222-010-9194-z>

Examples

if (FALSE) {

 #--- EXAMPLE 1 ------------------------------------------

 sig <- diag (2)
 cen <- rep (1, 2)
 x <- rbind(MASS::mvrnorm(108, cen * 0,   sig),
 	       MASS::mvrnorm(162, cen * 5,   sig * 6 - 2),
 	       MASS::mvrnorm(30, cen * 2.5, sig * 50))

 ctl <- ctlcurves(x, k = 1:4)
 ctl

   ##  ctl-curves 
 plot(ctl)  ##  --> selecting k = 2, alpha = 0.08

   ##  the selected model 
 plot(tclust(x, k = 2, alpha = 0.08, restr.fact = 7))

 #--- EXAMPLE 2 ------------------------------------------

 data(geyser2)
 ctl <- ctlcurves(geyser2, k = 1:5)
 ctl
 
   ##  ctl-curves 
 plot(ctl)  ##  --> selecting k = 3, alpha = 0.08

   ##  the selected model
 plot(tclust(geyser2, k = 3, alpha = 0.08, restr.fact = 5))


 #--- EXAMPLE 3 ------------------------------------------
 
 data(swissbank)
 ctl <- ctlcurves(swissbank, k = 1:5, alpha = seq (0, 0.3, by = 0.025))
 ctl
 
   ##  ctl-curves 
 plot(ctl)  ##  --> selecting k = 2, alpha = 0.1
 
   ##  the selected model
 plot(tclust(swissbank, k = 2, alpha = 0.1, restr.fact = 50))
 
}