Re: Collective Intelligence in Action; Implementing an SVM

[Paul Elschot <pa...@xs4all.nl>]

Op Wednesday 05 March 2008 09:10:48 schreef Isabel Drost:
> On Wednesday 05 March 2008, Lukas Vlcek wrote:
> > Based on what's inside section it
> > seems as it can be useful for anybody who would like to get more
> > familiar with Lucene/Nutch/WEKA/JDM (Java Data Mining API).
>
> I did have a look at it but did not read it through. At least the
> section about svms is pretty good for ml users - yet it did not
> explain exactly how to implement an svm.

You don't really have to implement an svm. For text purposes
this works nicely for me for small problems:

http://people.cs.uchicago.edu/~vikass/svmlin.html

I must say though that I have not tried it yet on larger a scale.
At the moment I'm toying (basteln, DE) with a system that can
feed svmlin (running as a java process) from the termvectors
of a small lucene index (450Kb).
Svmlin works as advertised, so far this problem seems to be
too easy.
But no cross validation yet, only classifying the test data.

Schoene Gruesse,
Paul Elschot