R help - Jul 2009 - stepwise variable selection method wanted

Hi List,

I am looking for a variable selection procedure with a forward-backward
selection method.
Firstly, it is meant to work with the cophenetic
correlation coefficient (CPCC) and intended to find the variable combination
with the
highest cophenetic correlation. Secondly, it is aimed at Gower metric with
wards method (though this could be easily extended) aimed at categorical data.

What I have so far is a function for backward selection that returns the
variables
deleted and associated CPCC.

My current approach is cumbersome and very slow when working with large data
sets (mostly
because of the proximity matrix calculation). There are also problems with using
only
backward selection, so a way of combining forward-backward would be much better.
I was hoping that someone has a better /faster selection procedure that can be
adapted to using the CPCC.

Below my backward selection function and example.

Thanks and cheers
Herry

################################################

require(cluster)

cophenCbw<-function(dta){
# cophenetic variable selection backward
if(!is.data.frame(dta)) {print("x must be a dataframe with variables as
columns, cases as rows")}
else if(ncol(dta) <3) {pring("input dataframe must have at least 3
columns")}
else {
#currently function only performs cophenC on gower with ward, but this can be
adjusted easily to other metrics/methods
require(cluster)
require(ade4)
dta->dta.sic
lhs<-dta
for(j in 1:ncol(dta)){
 print(paste("round", j))
 as.data.frame(matrix(ncol=4, nrow=0))->testm

 for(i in 0:ncol(lhs)) {
  if(i == 0){
   daisy(lhs, metric="gower")->d.all
     agnes(lingoes(d.all),method="ward")->agnes.d.all
     cophenetic(agnes.d.all)->d1
     cor(d1,d.all)->cc
     testm<-data.frame(varID=0,cophenC=round(cc,3),varsdel=NA,round=0)
  }
  else {
     daisy(lhs[,-i], metric="gower")->d.all
     agnes(lingoes(d.all),method="ward")->agnes.d.all
     cophenetic(agnes.d.all)->d1
     cor(d1,d.all)->cc
    
testm<-rbind(testm,data.frame(varID=i,cophenC=round(cc,3),varsdel=colnames(dta)[i],round=j))
     #print(paste("var", i, "out
of",ncol(lhs),"nrows",
nrow(lhs),"rowsInTestm:",nrow(testm)))
   }
 }
 if(j == 1) {
  testm[testm[,2] ==  min(testm[,2]),][1,]->varsdel    #use only the first if
there are several
   vars2del<-varsdel[j,3]
  lhs<-dta[,-which(colnames(dta) %in% vars2del)]
  
print(paste("var2delete",varsdel[j,1],varsdel[j,3],"cophenC=",varsdel[j,2],"rowsInTestm:",nrow(testm)))


 }   # put exclusion variable into record
 else {
  rbind(varsdel,testm[testm[,2] ==  min(testm[,2]),])->varsdel
   vars2del<-rbind(vars2del,varsdel[j,3])
  lhs<-dta[,-which(colnames(dta) %in% vars2del)]
  
print(paste("var2delete",varsdel[j,1],varsdel[j,3],"cophenC=",varsdel[j,2],"rowsInTestm:",nrow(testm)))
  }
 if(is.na(varsdel[j,3])) break

  }
  
 }
return(varsdel)
}

cophenCbw(plantTraits)


########################################################

Dear Herry,

I would like to know if you found an answer elsewhere to your question.
I'm trying to get information around the nodes of a CA (daisy() followed
by agnes()) made on plant trait using the gower metric for taxonomic
purpuse. I'm not an expert in statistic but I understood that your way
might be the only way to get the meaning of the structure of a such CA,
even if quite slow.
I already started to implement your function to use the weight and type
(logr, asymm, symm and ordr) dynamically. But maybe some easier and less
CPU cost way does exist now... somewhere...
I'm having bad time with R but you might help.

 Thanks for you answer.