coef_spicefp: coef_spicefp

In SpiceFP: Sparse Method to Identify Joint Effects of Functional Predictors

coef_spicefp

Description

This function allows to obtain the coefficients of a model (involving a candidate matrix and 2 regularization parameters). There are two possible options to use this function: 1/ by minimizing an information criterion and selecting a number of model (option by default), or 2/ directly by providing the parameters of the model(s) that the user wishes to reconstruct.

Usage

coef_spicefp(
  spicefp.result,
  iter_,
  criterion = "AIC_",
  nmodels = 1,
  model.parameters = NULL,
  dim.finemesh = NULL,
  ncores = parallel::detectCores() - 1,
  write.external.file = TRUE
)

Arguments

spicefp.result	List. Outputs of the spicefp function.
iter_	integer. number of the iteration of interest.
criterion	character. One of "AIC_", "BIC_", "Cp_". Can be NULL, "AIC_" by default. If specified, nmodels must also be provided.
nmodels	integer. Equal to 1 by default. Represents the number of best models, according to the information criterion used. Should be NULL if criterion = NULL.
model.parameters	data.frame. NULL by default. One or more rows contained in the file where the model statistics were stored. Be careful to use the file related to the selected iteration. Names used in model.parameters shoud be the same in the file.
dim.finemesh	numeric vector of length 2 or 3. This vector informs about the dimension of the fine-mesh arrays (or matrices).
ncores	numbers of cores that will be used for parallel computation. By default, it is equal to detectCores()-1.
write.external.file	logical. indicates whether the result table related to each iteration has been written as a file (txt) in your working directory. This argument must be equal to the argument with the same name in the spicefp function.

Details

By providing criterion and nmodels, the function returns the coefficients of the nmodels best models chosen by the selected information criterion. When model.parameters is instead provided, it returns the coefficients of the models described on each row of the data.frame.

Value

Returns a list of 2 elements:

Model.parameters

data.frame where each row contains statistics related to the models of interest. Same as input if model.parameters is provided.

coef.list

List of length nmodels or the number of rows in Model.parameters. Each element of this list contains the model results as provided by the genlasso package, its coefficients without and with NA, a fine-mesh array with the coefficients, and the estimation of X \beta. Coefficients with NA are coefficient vector where the coefficient value of never-observed joint modalities is NA.

Examples

##linbreaks: a function allowing to obtain equidistant breaks
linbreaks<-function(x,n){
       sort(round(seq(trunc(min(x)),
                ceiling(max(x)+0.001),
                length.out =unlist(n)+1),
            1)
           )
}
# In this example, we will evaluate 2 candidates with 14 temperature
# classes and 15 irradiance classes. The irradiance breaks are obtained
# according to a log scale (logbreaks function) with different alpha
# parameters for each candidate (0.005, 0.01).
## Data and inputs
tpr.nclass=14
irdc.nclass=15
irdc.alpha=c(0.005, 0.01)
p2<-expand.grid(tpr.nclass, irdc.alpha, irdc.nclass)
parlist.tpr<-split(p2[,1], seq(nrow(p2)))
parlist.irdc<-split(p2[,2:3], seq(nrow(p2)))
parlist.irdc<-lapply(
     parlist.irdc,function(x){
     list(x[[1]],x[[2]])}
)
m.irdc <- as.matrix(Irradiance[,-c(1)])
m.tpr <- as.matrix(Temperature[,-c(1)])

# For the constructed models, only two regularization parameter ratios
# penratios=c(1/25,5) is used. In a real case, we will have to evaluate
# more candidates and regularization parameters ratio.
ex_sp<-spicefp(y=FerariIndex_Difference$fi_dif,
              fp1=m.irdc,
              fp2=m.tpr,
              fun1=logbreaks,
              fun2=linbreaks,
              parlists=list(parlist.irdc,
                            parlist.tpr),
              penratios=c(1/25,5),
              appropriate.df=NULL,
              nknots = 100,
              ncores =2,
              write.external.file = FALSE)

# coef_spicefp
## coefficients based on the parameters of the model
## focus on model selected by Mallows's Cp at iteration 1

start_time_spc <- Sys.time()
results.eval.iter1<-ex_sp$Evaluations[[1]]$Evaluation.results$evaluation.result
c.mdl <- coef_spicefp(ex_sp, iter_=1,
                      criterion =NULL,
                      nmodels=NULL,
  model.parameters=results.eval.iter1[which.min(results.eval.iter1$Cp_),],
                      ncores = 1,
                      write.external.file =FALSE)

g1<-c.mdl$coef.list$'231'$Candidate.coef.NA.finemeshed
g1.x<-as.numeric(rownames(g1))
g1.y<-as.numeric(colnames(g1))
duration_spc <- Sys.time() - start_time_spc

#library(fields)
#plot(c(10,2000),c(15,45),type= "n", axes = FALSE,
#     xlab = "Irradiance (mmol/m2/s - Logarithmic scale)",
#     ylab = "Temperature (deg C)",log = "x")
#rect(min(g1.x),min(g1.y),max(g1.x),max(g1.y), col="black", border=NA)
#image.plot(g1.x,g1.y,g1, horizontal = FALSE,
#           col=designer.colors(64, c("blue","white")),
#           add = TRUE)
#axis(1) ; axis(2)

## Let's visualize the same model from other arguments of coef_spicefp
c.crit <- coef_spicefp(ex_sp, iter_=1,
                       criterion ="Cp_",nmodels=1,
                       ncores = 1,
                       write.external.file =FALSE)
g2<-c.crit$coef.list$'231'$Candidate.coef.NA.finemeshed
g2.x<-as.numeric(rownames(g2))
g2.y<-as.numeric(colnames(g2))
#plot(c(10,2000),c(15,45),type= "n", axes = FALSE,
#     xlab = "Irradiance (mmol/m2/s - Logarithmic scale)",
#     ylab = "Temperature (deg C)",log = "x")
#rect(min(g2.x),min(g2.y),max(g2.x),max(g2.y), col="black", border=NA)
#image.plot(g2.x,g2.y,g2, horizontal = FALSE,
#           col=designer.colors(64, c("blue","white")),
#           add = TRUE)
#axis(1) ; axis(2)
closeAllConnections()

SpiceFP documentation built on June 7, 2023, 5:55 p.m.