R help - Jul 2007 - RWeka control parameters classifiers interface

Hello,

  I have some trouble in achieving the desired parametrisation
  for the weka classifier functions, using the package RWeka.

  The problem is, that the functions
  result=classifier(formula, data, subset, na.action, control =
Weka_control(mycontrol))
  do not seem to be manipulated by the mycontrol- arguments

  Perhaps this should be resepected via the handlers- argument ,
  but the documentation in this regard is rather sparse. 

# ------------- Examples

file     
=system.file("arff","iris.arff",package="RWeka")
data      =read.arff(file=file)     		
rownames(data)=1:nrow(data)           	
colnames(data)[ncol(data)]  ="class"  	

library(RWeka)

# Example: no parameter influence
mySMO
=make_Weka_classifier(name="weka/classifiers/functions/SMO",class=NULL,handlers=list());
# Using control =Weka_control()
m1
=mySMO(formula=class~.,data=data[,],control=Weka_control(K="weka.classifiers.functions.supportVector.PolyKernel",E=2))
m2
=mySMO(formula=class~.,data=data[,],control=Weka_control(K="weka.classifiers.functions.supportVector.PolyKernel",E=3))
m3
=mySMO(formula=class~.,data=data[,],control=c("K","weka.classifiers.functions.supportVector.PolyKernel","E",3))
# Using predefinded interface, does not work 
x1 
=SMO(formula=class~.,data=data[,],control=Weka_control(K="weka.classifiers.functions.supportVector.PolyKernel",E=2))
x2 
=SMO(formula=class~.,data=data[,],control=Weka_control(K="weka.classifiers.functions.supportVector.PolyKernel",E=3))
m1$call   
m2$call   
m3$call		
x1$call
x2$call
# no differences:
m1 
m2 
m3
x1
x2

Any suggestions?
Many thanks
Bjoern

many thanks
bjoern

On Wed, 11 Jul 2007 strinz at freenet.de wrote:
>   The problem is, that the functions
>   result=classifier(formula, data, subset, na.action, control =
Weka_control(mycontrol))
>   do not seem to be manipulated by the mycontrol- arguments
Yes, they are...not all parameter changes have always an effect on the
specified learner.
>   Perhaps this should be resepected via the handlers- argument ,
>   but the documentation in this regard is rather sparse.
Handlers are not needed here.

Re: sparse docs. In case you have not seen that paper already, there is a
technical report on the ideas behind RWeka:
  http://epub.wu-wien.ac.at/dyn/openURL?id=oai:epub.wu-wien.ac.at:epub-wu-01_ba6

Re: SMO. Compare

m1 <- SMO(Species ~ ., data = iris)
m2 <- SMO(Species ~ ., data = iris, control = Weka_control(
  K = "weka.classifiers.functions.supportVector.RBFKernel"))

which yield different results so the Weka_control() works.

The same happens if you register the mySMO() interface yourself. I'm not
sure why the "E = ..." argument has no influence on the SMO, please
check
the Weka docs for this particular learner.

Best,
Z

Hi,

  many thanks for the answer.
  It ist true, that for example

  m1 <- SMO(Species ~ ., data = iris, control = Weka_control( K =
"weka.classifiers.functions.supportVector.PolyKernel"))
  m2 <- SMO(Species ~ ., data = iris, control = Weka_control( K =
"weka.classifiers.functions.supportVector.RBFKernel"))
  deliver different results
  
  but
  m3 <- SMO(Species ~ ., data = iris, control = Weka_control( K =
"weka.classifiers.functions.supportVector.PolyKernel",E=2))
  m4 <- SMO(Species ~ ., data = iris, control = Weka_control( K =
"weka.classifiers.functions.supportVector.RBFKernel"),G=0.2)

  m3 does not differ from m1 (from the point of view of the setup, irrespective
of the data!)
  m4 does not differ from m2 (from the point of view of the setup, irrespective
of the data!)

  which can be seen, when looking at the results:
  m1 # (Linear Kernel, okay )
  m2 # (RBF Kernel, okay)
  m3  # still uses a linear kernel, but should be a <x,y>^2 kernel
  m4 # G is ignored, resulting in m2


Thx
Bjoern




----- original Nachricht --------

Betreff: Re: [R] RWeka control parameters classifiers interface
Gesendet: Mi 11 Jul 2007 14:42:10 CEST
Von: "Achim Zeileis"<Achim.Zeileis at wu-wien.ac.at>
> On Wed, 11 Jul 2007 strinz at freenet.de wrote:
> 
> >   The problem is, that the functions
> >   result=classifier(formula, data, subset, na.action, control >
Weka_control(mycontrol))
> >   do not seem to be manipulated by the mycontrol- arguments
> 
> Yes, they are...not all parameter changes have always an effect on the
> specified learner.
> 
> >   Perhaps this should be resepected via the handlers- argument ,
> >   but the documentation in this regard is rather sparse.
> 
> Handlers are not needed here.
> 
> Re: sparse docs. In case you have not seen that paper already, there is a
> technical report on the ideas behind RWeka:
>  
>
http://epub.wu-wien.ac.at/dyn/openURL?id=oai:epub.wu-wien.ac.at:epub-wu-01_b
> a6
> 
> Re: SMO. Compare
> 
> m1 <- SMO(Species ~ ., data = iris)
> m2 <- SMO(Species ~ ., data = iris, control = Weka_control(
>   K = "weka.classifiers.functions.supportVector.RBFKernel"))
> 
> which yield different results so the Weka_control() works.
> 
> The same happens if you register the mySMO() interface yourself. I'm
not
> sure why the "E = ..." argument has no influence on the SMO,
please check
> the Weka docs for this particular learner.
> 
> Best,
> Z
> 
> 
> 
--- original Nachricht Ende ----