R help - Feb 2006 - getting probabilities from SVM

I am using SVM to classify categorical data and I would like the
probabilities instead of the classification.  ?predict.svm says that its
only enabled when you train the model with it enabled, so I did that, but it
didn't work.  I can't even get it to work with iris.  The help file
shows
that probability = TRUE when training the model, but doesn't show an
example.  Then I try to predict with probabilities, I still only get
classifications back.  Anyone get this to work and can help me out?


Thanks,

Roger

attach(iris)

     # alternatively the traditional interface:
     x <- subset(iris, select = -Species)
     y <- Species
     model <- svm(x, y, probability = TRUE)
     pred <- predict(model, x, probability = TRUE)

	[[alternative HTML version deleted]]

Finally figured it out.  You have to extract it from the attributes.
Tricky.  Thanks anyway.
> attr(pred, "prob")[1:10,]

On 2/16/06, roger bos <roger.bos@gmail.com> wrote:
>
> I am using SVM to classify categorical data and I would like the
> probabilities instead of the classification.  ?predict.svm says that its
> only enabled when you train the model with it enabled, so I did that, but
it
> didn't work.  I can't even get it to work with iris.  The help file
shows
> that probability = TRUE when training the model, but doesn't show an
> example.  Then I try to predict with probabilities, I still only get
> classifications back.  Anyone get this to work and can help me out?
>
>
> Thanks,
>
> Roger
>
> attach(iris)
>
>      # alternatively the traditional interface:
>      x <- subset(iris, select = -Species)
>      y <- Species
>      model <- svm(x, y, probability = TRUE)
>      pred <- predict(model, x, probability = TRUE)
>
	[[alternative HTML version deleted]]

> Finally figured it out.  You have to extract it from the attributes.
> Tricky.  Thanks anyway.
> attr(pred, "prob")[1:10,]
Correct. 

Just for the records, the rationale behind this `tricky' design:

In addition to probabilites, predict.svm() (more precisely: libsvm) can also
compute the decision values. Common ways to handle `polymorph' prediction
types are, e.g, using a `type' argument in the predict() function, or to
return all variants in one list object. With a `type' argument, you need
several calls to predict() if you need, say, hard predictions _and_ the
probabilities. On the other hand, the probability and decision values features
were added to libsvm only when svm() in e1071 had already been around for a
while, so returning a list instead of a vector would have broken a lot of code.
So I decided to keep the `standard' predict behavior and to `hide'
special predictions in an attribute. If the latter had been available from the
beginning, I probably would have used the `type' approach.

Cheers,
David


On 2/16/06, roger bos <roger.bos at gmail.com>
wrote:
>
> I am using SVM to classify categorical data and I would like the
> probabilities instead of the classification.  ?predict.svm says that its
> only enabled when you train the model with it enabled, so I did that, but
it
> didn't work.  I can't even get it to work with iris.  The help file
shows
> that probability = TRUE when training the model, but doesn't show an
> example.  Then I try to predict with probabilities, I still only get
> classifications back.  Anyone get this to work and can help me out?
-- 
Dr. David Meyer
Department of Information Systems and Operations

Vienna University of Economics and Business Administration
Augasse 2-6, A-1090 Wien, Austria, Europe
Fax: +43-1-313 36x746 
Tel: +43-1-313 36x4393
HP:  http://wi.wu-wien.ac.at/~meyer/