[pymvpa] Spam No.3 (Bug?!)

[Yaroslav Halchenko]

> Is it the case that we can use any SVM w/ multi-class data, and that  
> the SVM implementation will handle this case correctly, or do we have  
> to do pair-wise splits beforehand to reduce it to a two-class problem?

some SVM implementations (native libsvm backend, and libsvm and gmnp in
shogun) have built-in pair-wise + voting multiclass SVM implementation.
For the others indeed you would need to use MultiClassClassifier.
If you just want to try multiclass -- we recommend starting with SMLR --
it has built-in multiclass with 1-vs-rest kind of selection of the
decision value

> The reason I ask (and am concerned) is that I have written a project  
> which uses PyMVPA. 
that is great! ;-)

> Now, while the feature selection seems to work, students have reported  
> that they get the same classification error using a Linear C-SVM, even  
> when they drastically change the C value (from 10^-7 all the way to  
> 10^10).
nothing can be better for bug-shooting than real user's use cases ;-)

it is really weird of what has happened and I guess knowing what exactly
SVM implementation was used and either data wasn't bogus (print
dataset.summary()) might help to reveal the problem... also I guess I
need to check the code later on to assure you that internally everything
works ok ;-)

also if performance is at chance -- then varying C wouldn't help much ;)

> Furthermore, the accuracy was the same for two different  
> random data subsets.
the same at chance? the same perfect? or the same random number? were
decision values the same?

> This sounds bad to me, but I'm not exactly sure  
> why, and I thought that the two vs. many classes issue might be a  
> source of problems.
many things could be the source... in such cases make them both run it
with

MVPA_DEBUG=.* your_analysis 1>analysisX_debug.log 2>&1

and send me both of those _debug.log files -- may be they could
highlight the difference?

> Any help would be greatly appreciated!

> Thanks,
> James.

> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/mailman/listinfo/pkg-exppsy-pymvpa


-- 
Yaroslav Halchenko
Research Assistant, Psychology Department, Rutgers-Newark
Student  Ph.D. @ CS Dept. NJIT
Office: (973) 353-1412 | FWD: 82823 | Fax: (973) 353-1171
        101 Warren Str, Smith Hall, Rm 4-105, Newark NJ 07102
WWW:     http://www.linkedin.com/in/yarik