On Mon, Jun 17, 2013 at 10:27 PM, Lars Buitinck <[email protected]> wrote:
> In pairwise L2R, loss functions are defined on pairs of samples, but
> only for those pairs that pertain to the same query (since ranking is
> done per query). So, you need the query information with the samples.
>
Actually, if I am not mistaken, the query grouping is not a requirement but
just an heuristic to speed up training.
Technically, samples that do not belong to the same query should be ranked
lower than those
that do. However, considering all possible pairs would be computationally
too expensive. Therefore,
learning algorithms typically only enforce that samples within the same
query group are properly ranked.
My 2 yens,
Mathieu
------------------------------------------------------------------------------
This SF.net email is sponsored by Windows:
Build for Windows Store.
http://p.sf.net/sfu/windows-dev2dev
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
