[R] tune an support vector machine

[Uwe Bohne]

   Thank you very much,

   your proposal is one practical way to check for significant features.
   I tried to check for all combination in a loop, but unfortunately there is a
   problem with NA values.
   Maybe anybody has an idea.

   This is my expansion of the former code:

   namen<-expand.grid(c("weight",NA),
   c("height",NA),c("width",NA),c("volume",NA), stringsAsFactors=FALSE)
   namen2<-as.data.frame(namen)
   for(i in 1:nrow(namen2)){
     assign(paste("a", i, sep = ""), namen2[i,])
   }

   This generates vectors containing the features.
   If i pick one of them i can produce a formula that i can use for svm tuning.
   For example

   a7
   a7q<-t(as.data.frame(a7[!is.na(a7)]))
   a7q
   a7f<-as.formula(paste("type~",paste(a7q,collapse="+")))
   a7f

   and

   svmtune_a7=tune.svm(a7f,  data=train,  kernel="radial", cost=2^(-2:5),
   gamma=2^(-2:1),cross=10)

   works as desired.
   So my key idea was to tune SVM with every possibel "a...f" formula and
   choose  the  best one according to the best performance measure in the
   summary.
   Unfortunately I just have problems to make it in a loop.
   I tried

   for(iin1:nrow(namen2)){paste("a",i,"q",sep="")<-t(as.data.frame(paste("a",
   i,"[!is.na(a",i,")]", sep="")))}

   and produced error. Probably i didnt paste correctly.
   Any ideas?
   Thanks a lot!
   Uwe
   Gesendet: Samstag, 07. Dezember 2013 um 08:26 Uhr
   Von: "Wuming Gong" <gongx030 at umn.edu>
   An: "Uwe Bohne" <balu555 at gmx.de>
   Cc: "r-help mailinglist" <r-help at r-project.org>
   Betreff: Re: [R] tune an support vector machine
   Hi Uwe,

   It looks SVM in e1071 and Kernlab does not support feature selection, but
   you can take a look at package penalizedSVM
   ([1]http://cran.r-project.org/web/packages/penalizedSVM/penalizedSVM.pdf).

   Or you can implement a SVM-RFE
   ([2]http://axon.cs.byu.edu/Dan/778/papers/Feature%20Selection/guyon*.pdf)by
   the alpha values returned by svm() in e1071 or ksvm() in Kernlab.

   Wuming

   On Fri, Dec 6, 2013 at 7:06 AM, Uwe Bohne <[3]balu555 at gmx.de> wrote:

        Hej all,
        actually i try to tune a SVM in R and use the package "e1071" wich
     works
        pretty well.
        I do some gridsearch in the parameters and get the best possible
     parameters
        for classification.
        Here is my sample code
        type<-sample(c(-1,1) , 20, replace = TRUE )
        weight<-sample(c(20:50),20, replace=TRUE)
        height<-sample(c(100:200),20, replace=TRUE)
        width<-sample(c(30:50),20,replace=TRUE)
        volume<-sample(c(1000:5000),20,replace=TRUE)
        data<-cbind(type,weight,height,width,volume)
        train<-as.data.frame(data)
        library("e1071")
        features <- c("weight","height","width","volume")
        (formula<-as.formula(paste("type ~ ", paste(features, collapse= "+"))))
        svmtune=tune.svm(formula,  data=train, kernel="radial", cost=2^(-2:5),
        gamma=2^(-2:1),cross=10)
        summary(svmtune)
        My question is if there is a way to tune the features.
        So in other words - what i wanna do is to try all possible combinations
     of
        features : for example use only (volume) or use (weight, height) or use
        (height,volume,width) and so on for the SVM  and to get the best
     combination
        back.
        Best wishes
        Uwe
     ______________________________________________
     [4]R-help at r-project.org mailing list
     [5]https://stat.ethz.ch/mailman/listinfo/r-help
     PLEASE do read the posting guide
     [6]http://www.R-project.org/posting-guide.html
     and provide commented, minimal, self-contained, reproducible code.

References

   1. http://cran.r-project.org/web/packages/penalizedSVM/penalizedSVM.pdf
   2. http://axon.cs.byu.edu/Dan/778/papers/Feature%20Selection/guyon*.pdf
   3. file://localhost/tmp/balu555@gmx.de
   4. file://localhost/tmp/R-help@r-project.org
   5. https://stat.ethz.ch/mailman/listinfo/r-help
   6. http://www.R-project.org/posting-guide.html