[pymvpa] Fwd: new feature selection algorithm

Per B. Sederberg persed at princeton.edu
Tue Sep 30 23:56:48 UTC 2008

In case y'all didn't see this...

---------- Forwarded message ----------
From: Sam Gershman <sjgershm at princeton.edu>
Date: Tue, Sep 30, 2008 at 7:43 PM
Subject: new feature selection algorithm
To: compmemlist at princeton.edu

Hi all,

Some people might be interested in this:
It just came out in PNAS. The paper describes an algorithm for feature
selection called "higher criticism." It is designed for a particular
classification setting dubbed "rare/weak": where the fraction of
useful features is small and the useful features are each too weak to
be useful on their own. The idea is to look at the distribution of
feature statistics and use the deviation from an expected null
distribution to set the feature selection threshold. Although based in
frequentist statistics and therefore fundamentally flawed, it shows
some interesting behavior that could be highly advantageous to MVPA
applied to fMRI. Apart from its explicit designation for the R/W
setting (which is obviously apt for fMRI), it doesn't require tuning
by cross-validation and the threshold has low variance (which might be
good when looking for consistent feature sets).


More information about the Pkg-ExpPsy-PyMVPA mailing list