Hi Alex,
I agree it would be nice to see a multinomial logistic regreesion in the
scikit :). Any plans for it?
Is your warning about interpreting the weights just because in a penalized
model they're not best estimators? Mostly I just want to take a look to see
in general what's happening -- e.g. does it look like there is some amount
of clustering into blobs, or are my weights mixed more or less randomly
(I'm using L2 regularization, so they aren't sparse). Also this isn't full
brain data, just restricted within a few ROIs (~500 features per model). To
the extent that I might do inference on them, it would be taking each
subject's weights to a group level random-effects test against 0. But
mostly I'm just trying to get a feel for what's driving my classification
and don't care too much about the exact values of the weights. I figured I
might as well take advantage of PySurfer's ability to draw things in lots
of pretty colors :)
Thanks!
Michael
On Sun, Mar 4, 2012 at 11:57 PM, Alexandre Gramfort <
[email protected]> wrote:
> Hi Michael,
>
> my answer is more a warning than an answer.
>
> Beyond 2 classes the "best" would be to use something like L1/L2 mixed norm
> with a real multinomial loss. Unfortunately we don't have it in the scikit.
>
> Also looking at the weights of a sparse logistic or L2 model (logistic or
> SVM)
> working with full brain data can be dangerous when it comes to
> "interpretation".
> Basically I wouldn't do it if it's a plain logistic regression working
> with voxel
> based features. Now it depends on what you want to say and you might
> find a way, eg. using permutations or bootstrap, to assess some statistical
> significance.
>
> Alex
>
> On Mon, Mar 5, 2012 at 4:13 AM, Michael Waskom <[email protected]>
> wrote:
> > Hi all,
> >
> > I have a LogisticRegression model I'm training in a 3-class scenario.
> I'd
> > like to examine the coefficients for the models. As the default for
> > LogisitcRegression is to do one-vs-all classification, my clf.coef_
> array is
> > shape 3 x nfeat.
> >
> > My question is how to interpret the sign of the coefficients. I take
> each
> > nfeat vector to be the coefficients for the A vs all, B vs all, C vs all
> > models. In this case, are positive coefficients in the first nfeat
> vector
> > those weighing the classification towards "A" and the negative
> coefficients
> > those weighing the model towads "all"? Or is it the other way around?
> >
> > Cheers,
> > Michael
> >
> >
> ------------------------------------------------------------------------------
> > Try before you buy = See our experts in action!
> > The most comprehensive online learning library for Microsoft developers
> > is just $99.99! Visual Studio, SharePoint, SQL - plus HTML5, CSS3, MVC3,
> > Metro Style Apps, more. Free future releases when you subscribe now!
> > http://p.sf.net/sfu/learndevnow-dev2
> > _______________________________________________
> > Scikit-learn-general mailing list
> > [email protected]
> > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
> >
>
>
> ------------------------------------------------------------------------------
> Try before you buy = See our experts in action!
> The most comprehensive online learning library for Microsoft developers
> is just $99.99! Visual Studio, SharePoint, SQL - plus HTML5, CSS3, MVC3,
> Metro Style Apps, more. Free future releases when you subscribe now!
> http://p.sf.net/sfu/learndevnow-dev2
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
------------------------------------------------------------------------------
Try before you buy = See our experts in action!
The most comprehensive online learning library for Microsoft developers
is just $99.99! Visual Studio, SharePoint, SQL - plus HTML5, CSS3, MVC3,
Metro Style Apps, more. Free future releases when you subscribe now!
http://p.sf.net/sfu/learndevnow-dev2
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
