[pymvpa] SensitivityBasedFeatureSelection for LinearCSVMC

[Yaroslav Halchenko]

On Sat, 17 Dec 2011, patrik andersson wrote:
> I have a basic question about the sensitivity analysis for the
> LinearCSVMC classifier.
> Namely: What does it do?
* for linear SVM it just outputs coefficients of the separating
  hyperplane
* if classifier was trained already it shouldn't retrain it

> Simple hey?! :)
yey

> 			   clfr.get_sensitivity_analyzer(postproc=maxofabs_sample()),
> 			   >...<
> What comes out of the  clfr.get_sensitivity_analyzer function?

the function which "extracts" hyperplane coefficients from SVM

> Is there some iteration or data folding going on inside?
for regular SVM -- nope.

> What happens if fsel is called with multiclass data?

that is what maxofabs_sample there for -- across all pairs of
classifiers (does all pair-wise + max vote) for each feature would
select maximal abs coefficient... somewhat ad-hoc and probably
suboptimal since coefficients are not unit-normed and would depend on
the margin width, so better would probably be first to norm all the
sensitives and then select max, e.g.

postproc=ChainMapper([ FxMapper('features', l2_normed),
                       maxofabs_sample() ] ),

> Answer this and I wish you an even happier x-mas ;)

;-) Merry X-mas to you too
-- 
=------------------------------------------------------------------=
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic