There was some discussion along these lines last year, but I don't think 
anyone has worked on it yet.  Scikit-learn doesn't currently have the 
ability to do manifold learning from a precomputed distance matrix, but 
it could be extended to that pretty easily.

What it would take would be to modify the various manifold learning 
classes to optionally take a precomputed distance matrix rather than a 
set of data.  There's some precedent for this in other algorithms (e.g. 
passing a precomputed Gram matrix in SVM/SVC).  It would be pretty 
straightforward to implement for isomap.  The LLE-based methods would 
take a bit more thought, but there's some info on how to do this in the 
Roweis & Saul paper.

I hope that helps - if you want to start a PR on this topic I'd 
certainly be available to give input along the way
   Jake

Anthony Bak wrote:
> I'd like to use isomap (and other manifold learning techniques) with
> abstract metric spaces (and perhaps more generally similarity and
> dissimilarity matricies - but we can put that aside for the moment).
> It looks to me like isomap assumes points are described by points in
> R^N or some data structure (such as a KD-Tree) built from such points.
>
> Q1: Can I use the version of isomap in sklearn with abstract metric spaces?
>
> I assumed that I could not based on a quick reading of the
> documentation six months or so ago and I wrote a python implementation
> (Based on the original Tenenbaum Matlab implementation).
>
> Q2: If the answer to Q1 is "no", how do I go about getting this more
> general isomap into the sklearn code?
>
> Do I need to make a case for handeling non-embedded data or are the
> advantages obvious to everyone?
>
> Thanks.
>
> ------------------------------------------------------------------------------
> For Developers, A Lot Can Happen In A Second.
> Boundary is the first to Know...and Tell You.
> Monitor Your Applications in Ultra-Fine Resolution. Try it FREE!
> http://p.sf.net/sfu/Boundary-d2dvs2
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>   

------------------------------------------------------------------------------
For Developers, A Lot Can Happen In A Second.
Boundary is the first to Know...and Tell You.
Monitor Your Applications in Ultra-Fine Resolution. Try it FREE!
http://p.sf.net/sfu/Boundary-d2dvs2
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to