Satra,
I don't think scikit-learn currently has the functionality you mention, 
but it's something that should be a goal.  A while ago I opened an issue 
addressing part of this:
  https://github.com/scikit-learn/scikit-learn/issues/351
Though inclusion of more general distance metrics in the tree codes 
would be difficult, it would be fairly straightforward to extend the 
brute-force Neighbors Classifier to use the distance metrics available 
in sklearn.pairwise, as well as perhaps an arbitrary user-defined distance.
This is one of my goals for the coding sprint in December.

For the future, it might be interesting to think about how more general 
distance metrics could be included in the BallTree.  It's possible in 
theory, because the Ball Tree algorithm works for any metric satisfying 
the triangle inequality.  We could write a libary of c/cython distance 
metric functions and use function pointers within the BallTree 
evaluation to allow the user to specify which metric they'd like to use.
   Jake

Satrajit Ghosh wrote:
> hi all,
>
> does anybody or some blackhole in sklearn have an implementation of a 
> kneighbors classifier with a custom distance function?
>
> cheers,
>
> satra
>
> ------------------------------------------------------------------------
>
> ------------------------------------------------------------------------------
> RSA® Conference 2012
> Save $700 by Nov 18
> Register now
> http://p.sf.net/sfu/rsa-sfdev2dev1
> ------------------------------------------------------------------------
>
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>   

------------------------------------------------------------------------------
RSA® Conference 2012
Save $700 by Nov 18
Register now
http://p.sf.net/sfu/rsa-sfdev2dev1
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to