I agree, that does look odd.
Can you open an issue on the tracker?
It does look like a numeric instability I don't know where it comes
from. Even with much smaller sigma the same happens.
Andy
On 01/16/2015 03:21 PM, Sylvain Takerkart wrote:
Dear all,
I'm trying to use the instance weighting capability of the SVC class,
and I encountered some weird behavior: when the C parameter is chosen
very small and the weights are not very large, weighted SVM yields
constant prediction...
Here is a piece of code that demonstrates this
https://dl.dropboxusercontent.com/u/2829280/wsvm_problem.py
What this does:
1. create a dummy dataset for binary classification, with a training
and a testing dataset
2. create random weights that are very very very close to one (so
close that they should not influence what follows)
3. run unweighted and weighted SVM (hereafter SVM and wSVM) with
decreasing values of C
We expect SVM and wSVM to yield exactly the same predictions because
the instance weights are ridiculously close to one. This is exactly
what happens for large to small-ish values of C; but at some point
when C gets smaller, wSVM yields constant predictions (either all
zeros, or all ones) while SVM still behaves normally...
Has anybody any clue about what's happening here? This looks like a
bug, with maybe some rounding problem somewhere???
Thanks for your help!
--
Sylvain Takerkart
Institut des Neurosciences de la Timone (INT)
UMR 7289 CNRS-AMU
Marseille, France
tél: +33 (0)4 91 324 007
http://www.int.univ-amu.fr/_TAKERKART-Sylvain_?lang=en
------------------------------------------------------------------------------
New Year. New Location. New Benefits. New Data Center in Ashburn, VA.
GigeNET is offering a free month of service with a new server in Ashburn.
Choose from 2 high performing configs, both with 100TB of bandwidth.
Higher redundancy.Lower latency.Increased capacity.Completely compliant.
http://p.sf.net/sfu/gigenet
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
------------------------------------------------------------------------------
New Year. New Location. New Benefits. New Data Center in Ashburn, VA.
GigeNET is offering a free month of service with a new server in Ashburn.
Choose from 2 high performing configs, both with 100TB of bandwidth.
Higher redundancy.Lower latency.Increased capacity.Completely compliant.
http://p.sf.net/sfu/gigenet
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
