[Xapian-devel] Learning to Rank : GSoC 2012

[Rishabh Mehrotra]

Hi Ashish,

As your doubt related to the algorithms is a general one, I would like to
try addressing it. Ranknet is a pairwise approach while ListNet is a
listwise approach to ranking, so Listnet's advantages over Ranknet would be
same as what other Listwise algorithms have over Pairwise ones.

The listwise approach addresses the ranking problem in the following way.
In learning, it takes ranked lists of objects as instances and trains a
ranking function through the minimization of a listwise loss function
defined on the predicted list and the ground truth list. The
listwise approach captures the ranking problems in a conceptually more
natural way than pairwise, apart from the computational advantages(I am of
sure of the specific here).

For your other doubt on the Adarank: the inherent advantage of
Adarank(build on the Adaboost concept) is that it minimizes a loss
function directly defined on the performance measures with respect to "the
training data". It re-weighs the training instances while constructing weak
learners and in the end forms an ensemble of these weak-learners aiming for
the total performance to be "boosted". In the case of linear regression, we
don't give different weights to different training tuples and build an
ensemble in the end: we work with just one model.
You could refer to the original paper here:
[link<http://research.microsoft.com/en-us/people/hangli/xu-sigir07.pdf>
].

Hope it helps! Do let me know if I have written anything incorrect above. :)

Refards,
Rishabh.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.xapian.org/pipermail/xapian-devel/attachments/20120401/13f7540f/attachment.htm>