Funclustering (version 1.0.2)

funclust: funclust, clustering multivariate functional data

Description

This function run clustering algorithm for multivariate functional data.

Usage

funclust(fd, K, thd = 0.05, increaseDimension = FALSE,
    hard = FALSE, fixedDimension = integer(0), nbInit = 5,
    nbIterInit = 20, nbIteration = 200, epsilon = 1e-05)

Arguments

fd

in the univariate case fd is an object from a class fd of fda package. Otherwise, in the multivariate case, fd is a list of fd objects (fd=list(fd1,fd2,..)).

K

the number of clusters.

thd

the threshold in the Cattell scree test used to select the numbers of principal components retained for the approximation of the probability density of the functional data.

increaseDimension

A logical parameter. If FALSE, the numbers of principal components are selected at each step of the algorithm according to the Cattell scree test. If TRUE, only an increase of the numbers of principal components are allowed.

hard

A logical parameter. If TRUE, the algorithm is randomly initialized with "hard" cluster membership (each curves is randomly assigned to one of the clusters). if FALSE, "soft" cluster membership are used (the cluster membership probabilities are randomly choosen in the K-simplex).

fixedDimension

A vector of size K which contains the numbers of principal components in the case of fixed numbers of principal components.

epsilon

The stoping criterion for the EM-like algorithm: the algorithm is stopped if the difference between two successive loglikelihood is less than epsilon.

nbInit

The number of small-EM used to determine the intialization of the main EM-like algorithm.

nbIterInit

The maximum number of iterations for each small-EM.

nbIteration

The maximum number of iteration in the main EM-like algorithm.

Value

a list containing: tik: conditional probabilities of cluster membership for each observation cls: the estimated partition, proportions: the mixing proportions, loglikelihood: the final value of the approximated likelihood, loglikTotal: the values of the approximated likelihood along the EM-like algorithm (not necessarily increasing), aic, bic, icl: model selection criteria, dimensions: number of principal components, used in the functional data probability density approximation, selected for each clusters, dimTotal: values of the number of principal components along the EM-like algorithm, V: principal components variances per cluster

Details

There is multiple kind of running the function funclust. The first one is to run the function with fixed dimensions among all iteration of the algorithm. (parameter fixedDimenstion). fixedDimension must be an integer vector of size K (number of cluster). If the user gives a not integer value in fixedDimension, then this value will be convert automatically to an integer one for examplethe, the algorithm will run with a dimension 2 instead of 2.2 given by user. The second one is to run it with a dimensions variying according to the results of the scree test (parameter increaseDimension). If increaseDimension = true, then the dimensions will be constraint to only increase between to consecutuves iterations of the algorithm. else the values of the dimensions will be the results of the scree test.

Examples

# NOT RUN {
data(growth)
data=cbind(matrix(growth$hgtm,31,39),matrix(growth$hgtf,31,54));
t=growth$age;
splines <- create.bspline.basis(rangeval=c(1, max(t)), nbasis = 20,norder=4);
fd <- Data2fd(data, argvals=t, basisobj=splines);
# with varying dimensions (according to the results of the scree test)
res=funclust(fd,K=2)
summary(res)

# with fixed dimensions
res=funclust(fd,K=2,fixedDimension=c(2,2))


# Multivariate (deactivated by default to comply with CRAN policies)
# ---------  CanadianWeather (data from the package fda) --------
# CWtime<- 1:365
# CWrange<-c(1,365)
# CWbasis <- create.fourier.basis(CWrange, nbasis=65)
# harmaccelLfd <- vec2Lfd(c(0,(2*pi/365)^2,0), rangeval=CWrange)

# -- Build the curves ---
# temp=CanadianWeather$dailyAv[,,"Temperature.C"]
# CWfd1 <- smooth.basisPar(
# CWtime, CanadianWeather$dailyAv[,,"Temperature.C"],CWbasis,
# Lfdobj=harmaccelLfd, lambda=1e-2)$fd
# precip=CanadianWeather$dailyAv[,,"Precipitation.mm"]
# CWfd2 <- smooth.basisPar(
# CWtime, CanadianWeather$dailyAv[,,"Precipitation.mm"],CWbasis,
# Lfdobj=harmaccelLfd, lambda=1e-2)$fd

# CWfd=list(CWfd1,CWfd2)

# res=funclust(CWfd,K=2)
# }