[pymvpa] Q: Multi-class SVM which returns p-vals for each of the classes?

[Rajeev Raizada]

Dear PyMVPA folks,

Does the toolbox include an SVM version
which returns the p-vals for each class,
in a multiclass classification?

I know that libsvm returns probabilities,
and I have been trying this via its Matlab interface.
However, rather than returning a nice list of N p-vals
for each of the N possible class-label assignments,
it returns the p-vals for all of the pairwise two-class sub-classifications.
In the case of 8 classes, there are 8*7/2 = 28 pairwise p-vals.
Although there is almost certainly a simple way
of transforming that collection of pairwise p-vals
into a simple one-pval-per-class output, I don't know what it is.

I sent a mail about that topic to the libsvm board,
but it hasn't been answered yet:
http://agbs.kyb.tuebingen.mpg.de/km/bb/showthread.php?tid=2104&pid=5606#pid5606

I know that PyMVPA includes wrappers around
lots of different SVM implementations,
so I was wondering if any of those implementations
offers a more user-friendly outputting of multiclass p-values.

In the mean-time, I can get p-values out of other
multi-class classifiers, such as SMLR.
But it would be nice to compare them against SVM p-values.

Many thanks,

Raj