Other publication
Efficient optimization approaches for pairwise ranking losses
Authors: Antti Airola
Publication year: 2014
Journal: Dagstuhl Reports
Volume: 4
Issue: 3
First page : 8
Last page: 8
ISSN: 2192-5283
DOI: https://doi.org/10.4230/DagRep.4.3.1
Web address : http://drops.dagstuhl.de/opus/volltexte/2014/4550/
Abstract
Straightforward approaches to minimizing pairwise ranking losses on scored data lead to quadratic costs. We demonstrate, that for the special cases of pairwise hinge loss (RankSVM) and pairwise least-squares loss (RankRLS), better scaling can be achieved by modeling the preferences only implicitly using suitable data structures. Software implementations are available at http://staff.cs.utu.fi/~aatapa/software/RankSVM/(RankSVM) and https://github.com/aatapa/RLScore(RankRLS)
Straightforward approaches to minimizing pairwise ranking losses on scored data lead to quadratic costs. We demonstrate, that for the special cases of pairwise hinge loss (RankSVM) and pairwise least-squares loss (RankRLS), better scaling can be achieved by modeling the preferences only implicitly using suitable data structures. Software implementations are available at http://staff.cs.utu.fi/~aatapa/software/RankSVM/(RankSVM) and https://github.com/aatapa/RLScore(RankRLS)
UTU Research Portal