[Pkg-exppsy-maintainers] First round of SMLR working!

Per B. Sederberg persed at princeton.edu
Wed Feb 27 13:36:27 UTC 2008


Hi folks:

I just committed the first working version of sparse multinomial
logistic regression (SMLR).  I may be jumping ahead, but I think this
classifier will be revolutionizing the way we do business.

If you look at the example (and run it too) you'll see that when
starting with a dataset with 10000 features (but only 10
real/informative features, i.e., 9990 features of pure noise), it can
whittle them down to ~100 and get almost perfect classification.  If
you run the same code with a support vector, my guess is that it'll be
a chance (I'll add that into the example.)

We've started running this on whole-brains and it looks to be
performing feature selection and classification (both in a
simultaneous multivariate fashion) better than any other method we
have to date.

So, try it out!!!  I'll be adding unit tests and more examples and
trying it more on actual data.  We should probably make some
sensitivity analyzers out of it, too, because it can be used as a
feature-selection algorithm alone :)

Best,
Per

PS-> Don't be too annoyed with the speed of classification.  It's
loops by design, which can't be vectorized, but they will be very very
fast when I write that part in C.  Very very fast.



More information about the Pkg-exppsy-maintainers mailing list