vignettes/UseNbcomp.Rmd
UseNbcomp.RmdThis vignette is a short tutorial on the use of the core functions of
the bootPLS package.
These functions can also be used to reproduce the simulations and the figures of the book chapter Magnanensi et al. (2016) and of the two articles https://doi.org/10.1007/978-3-319-40643-5_18, Magnanensi et al. (2017) https://doi.org/10.1007/s11222-016-9651-4 and Magnanensi et al. (2021) https://doi.org/10.3389/fams.2021.693126.
Load and display the pinewood worm dataset.
library(bootPLS)
library(plsRglm)
data(pine, package = "plsRglm")
Xpine<-pine[,1:10]
ypine<-log(pine[,11])
pairs(pine)Michel Tenenhaus’ reported in his book, La régression PLS (1998) Technip, Paris, that most of the expert biologists claimed that this dataset features two latent variables, which is tantamount to the PLS model having two components.
Leave one out CV (K=nrow(pine)) one time
(NK=1).
bbb <- plsRglm::cv.plsR(log(x11)~.,data=pine,nt=6,K=nrow(pine),NK=1,verbose=FALSE)
plsRglm::cvtable(summary(bbb))Set up 6-fold CV (K=6), 100 times (NK=100),
and use random=TRUE to randomly create folds for repeated
CV.
Display the results of the cross-validation.
The criterion is recommended in that PLSR setting without missing data. A model with 1 component is selected by the cross-validation as displayed by the following figure. Hence the criterion (1 component) does not agree with the experts (2 components).
As for the CV Press criterion it is unable to point out a unique number of components.
The package features our bootstrap based algorithm to select the
number of components in plsR regression. It is implemented with the
nbcomp.bootplsR function.
set.seed(4619)
nbcomp.bootplsR(Y=ypine,X=Xpine,R =500)The verbose=FALSE option suppresses messages output
during the algorithm, which is useful to replicate the bootstrap
technique. To set up parallel computing, you can use the
parallel and the ncpus options.
set.seed(4619)
res_boot_rep <- replicate(20,nbcomp.bootplsR(Y=ypine,X=Xpine,R =500,verbose =FALSE,parallel = "multicore",ncpus = 2))It is easy to display the results with the barplot
function.
A model with two components should be selected using our bootstrap based algorithm to select the number of components. Hence the number of component selected with our algorithm agrees with what was stated by the experts.
The package also features our bootstrap based algorithm to select,
for a given
value, the number of components in spls regression. It is implemented
with the nbcomp.bootspls function.
nbcomp.bootspls(x=Xpine,y=ypine,eta=.5)A doParallel and foreach based parallel
computing version of the algorithm is implemented as the
nbcomp.bootspls.para function.
nbcomp.bootspls.para(x=Xpine,y=ypine,eta=.5)
nbcomp.bootspls.para(x=Xpine,y=ypine,eta=c(.2,.5))Pinewood worm data reloaded.
library(bootPLS)
library(plsRglm)
data(pine, package = "plsRglm")
Xpine<-pine[,1:10]
ypine<-log(pine[,11])
datasetpine <- cbind(ypine,Xpine)
coefs.plsR.adapt.ncomp(datasetpine,sample(1:nrow(datasetpine)))Replicate the results to get the bootstrap distributions of the selected number of components and the coefficients.
replicate(20,coefs.plsR.adapt.ncomp(datasetpine,sample(1:nrow(datasetpine))))Parallel computing support with the ncpus and
parallel="multicore" options.
coefs.plsR.adapt.ncomp(datasetpine,sample(1:nrow(datasetpine)),ncpus=2,parallel="multicore")
replicate(20,coefs.plsR.adapt.ncomp(datasetpine,sample(1:nrow(datasetpine)),ncpus=2,parallel="multicore"))Loading the data and creating the data frames.
data(aze_compl)
Xaze_compl<-aze_compl[,2:34]
yaze_compl<-aze_compl$y
dataset <- cbind(y=yaze_compl,Xaze_compl)Fitting a logistic PLS regression model with 10 components. You have to use the family option when fitting the plsRglm.
modplsglm <- plsRglm(y~.,data=dataset,10,modele="pls-glm-family",family="binomial")Perform the bootstrap based algorithm with the
nbcomp.bootplsRglm function. By default 250 resamplings are
carried out.
set.seed(4619)
aze_compl.bootYT <- suppressWarnings(nbcomp.bootplsRglm(modplsglm))Plotting the bootstrap distributions of the coefficients of the components.
plsRglm::boxplots.bootpls(aze_compl.bootYT)Computing the bootstrap based confidence intervals of the coefficients of the components.
plsRglm::confints.bootpls(aze_compl.bootYT)Computing the bootstrap based confidence intervals of the coefficients of the components.
suppressWarnings(plsRglm::plots.confints.bootpls(plsRglm::confints.bootpls(aze_compl.bootYT)))The package also features our bootstrap based algorithm to select,
for a given
value, the number of components in sgpls regression. It is implemented
in the nbcomp.bootsgpls.
set.seed(4619)
data(prostate, package="spls")
nbcomp.bootsgpls((prostate$x)[,1:30], prostate$y, R=250, eta=0.2, typeBCa = FALSE)A doParallel and foreach based parallel
computing version of the algorithm is implemented as the
nbcomp.bootspls.para function.
nbcomp.bootsgpls.para((prostate$x)[,1:30], prostate$y, R=250, eta=c(.2,.5), maxnt=10, typeBCa = FALSE)