flexmix (version 2.3-19)

stepFlexmix: Run FlexMix Repeatedly

Description

Runs flexmix repeatedly for different numbers of components and returns the maximum likelihood solution for each.

Usage

initFlexmix(..., k, init = list(), control = list(), nrep = 3L,
            verbose = TRUE, drop = TRUE, unique = FALSE)
initMethod(name = c("tol.em", "cem.em", "sem.em"),
           step1 = list(tolerance = 10^-2),
           step2 = list(), control = list(), nrep = 3L)
stepFlexmix(..., k = NULL, nrep = 3, verbose = TRUE, drop = TRUE,
            unique = FALSE)
# S4 method for stepFlexmix,missing
plot(x, y, what = c("AIC", "BIC", "ICL"),
  xlab = NULL, ylab = NULL, legend = "topright", ...)
# S4 method for stepFlexmix
getModel(object, which = "BIC")
# S4 method for stepFlexmix
unique(x, incomparables = FALSE, ...)

Value

An object of class "stepFlexmix" containing the best models with respect to the log likelihood for the different number of components in a slot if length(k)>1, else directly an object of class "flexmix".

If unique = FALSE, then the resulting object contains one model per element of k (which is the number of clusters the EM algorithm started with). If unique = TRUE, then the result is resorted according to the number of clusters contained in the fitted models (which may be less than the number with which the EM algorithm started), and only the maximum likelihood solution for each number of fitted clusters is kept. This operation can also be done manually by calling unique() on objects of class

"stepFlexmix".

Arguments

...

Passed to flexmix (or matplot in the plot method).

k

A vector of integers passed in turn to the k argument of flexmix.

init

An object of class "initMethod" or a named list where initMethod is called with it as arguments in addition to the control argument.

name

A character string indication which initialization strategy should be employed: short runs of EM followed by a long ("tol.em"), short runs of CEM followed by a long EM run ("cem.em"), short runs of SEM followed by a long EM run ("sem.em").

step1

A named list which combined with the control argument is coercable to a "FLXcontrol" object. This control setting is used for the short runs.

step2

A named list which combined with the control argument is coercable to a "FLXcontrol" object. This control setting is used for the long run.

control

A named list which combined with the step1 or the step2 argument is coercable to a "FLXcontrol" object.

nrep

For each value of k run flexmix nrep times and keep only the solution with maximum likelihood. If nrep is set for the long run, it is ignored, because the EM algorithm is deterministic using the best solution discovered in the short runs for initialization.

verbose

If TRUE, show progress information during computations.

drop

If TRUE and k is of length 1, then a single flexmix object is returned instead of a "stepFlexmix" object.

unique

If TRUE, then unique() is called on the result, see below.

x, object

An object of class "stepFlexmix".

y

Not used.

what

Character vector naming information criteria to plot. Functions of the same name must exist, which take a stepFlexmix object as input and return a numeric vector like AIC,stepFlexmix-method (see examples below).

xlab,ylab

Graphical parameters.

legend

If not FALSE and what contains more than 1 element, a legend is placed at the specified location, see legend for details.

which

Number of model to get. If character, interpreted as number of components or name of an information criterion.

incomparables

A vector of values that cannot be compared. Currently, FALSE is the only possible value, meaning that all values can be compared.

Author

Friedrich Leisch and Bettina Gruen

References

Friedrich Leisch. FlexMix: A general framework for finite mixture models and latent class regression in R. Journal of Statistical Software, 11(8), 2004. doi:10.18637/jss.v011.i08

Christophe Biernacki, Gilles Celeux and Gerard Govaert. Choosing starting values for the EM algorithm for getting the highest likelihood in multivariate Gaussian mixture models. Computational Statistics & Data Analysis, 41(3--4), 561--575, 2003.

Theresa Scharl, Bettina Gruen and Friedrch Leisch. Mixtures of regression models for time-course gene expression data: Evaluation of initialization and random effects. Bioinformatics, 26(3), 370--377, 2010.

Examples

data("Nclus", package = "flexmix")

## try 2 times for k = 4
set.seed(511)
ex1 <- initFlexmix(Nclus~1, k = 4, model = FLXMCmvnorm(diagonal = FALSE),
                   nrep = 2)
ex1

## now 2 times each for k = 2:5, specify control parameter
ex2 <- initFlexmix(Nclus~1, k = 2:5, model = FLXMCmvnorm(diagonal = FALSE),
                   control = list(minprior = 0), nrep = 2)
ex2
plot(ex2)

## get BIC values
BIC(ex2)

## get smallest model
getModel(ex2, which = 1)

## get model with 3 components
getModel(ex2, which = "3")

## get model with smallest ICL (here same as for AIC and BIC: true k = 4)
getModel(ex2, which = "ICL")

## now 1 time each for k = 2:5, with larger minimum prior
ex3 <- initFlexmix(Nclus~1, k = 2:5,
                   model = FLXMCmvnorm(diagonal = FALSE),
                   control = list(minprior = 0.1), nrep = 1)
ex3

## keep only maximum likelihood solution for each unique number of
## fitted clusters:
unique(ex3)