[pymvpa] Releasing a python wrapper around Deng's mRMR feature selection algorithm

[Rajeev Raizada]

On Wed, Jun 1, 2011 at 1:25 PM, Yaroslav Halchenko
<debian at onerussian.com> wrote:
> Thank you Brice for sharing -- have you tried/played with it on some
> neural data (fMRI, EEG, ...)?  Meanwhile let's also consult our local
> experts:
>
> Hi Raj,
> Only google knows that docstoc knows that you had some plans to
> investigate mRMR in terms of benefits/applicability of mRMR to fMRI
> data.  I wonder if you got a chance to play with it and what would be
> the summary?

I did play with it a little bit, but I didn't find any great results with it.
Like pretty much any other feature selection scheme,
it first tries to select the features, voxels in this case,
which are individually the most informative, i.e. a univariate measure.

mRMR's novel contribution is that it then tries to prune down that list
so that it doesn't end up selecting a whole bunch of features
which are all carrying the same information.

My guess is that this probably works quite well in genetics,
where SNPs which happen to sit next to each on a DNA strand
do not actually carry the same information.
However, in fMRI, neighbouring voxels do often carry similar info.
So, mRMR ends up trying to select spatially non-contiguous voxels.

That may or may not be a good thing, depending on what you want.
In my view, a difficulty of feature-selection is that it usually ends up
selecting widely varying salt-and-pepper patterns of scattered voxels
across all the different subjects, so it is difficult to know how to combine
those selected voxels across subjects.
mRMR tends to make that problem worse.

However, I didn't explore it nearly enough to be sure
that I had truly squeezed all the potential juice out of it.
I wouldn't be at all surprised if it could be interestingly useful for fMRI.
I just didn't personally get much out of it, during my brief exploration.

Raj