Perturbation based Large Margin Approach for Ranking

We consider the task of devising large-margin based surrogate losses for the learning to rank problem. In this learning to rank setting, the traditional hinge loss for structured outputs faces two main challenges: (a) the supervision consists of instances with multiple training labels instead of a single label per instance, and (b) the label space of the set of all permutations of items is very large, and less amenable to the usual dynamic programming based methods. The most natural way to deal with multiple labels leads, unfortunately, to a non-convex surrogate. We address this by first providing a general class of convex perturbation based surrogates as an extension of the large margin method. Our experiments demonstrate that a simple surrogate from this class performs better than other candidate large margin proposals for the learning to rank task.

• Type: Journal
• Year: 2012
• Where: JMLR
• Authors: Eunho Yang, Ambuj Tewari, Pradeep D. Ravikumar
• Keywords: Dynamic Programming, JMLR 2012, Label Space, Permutations, Programming Languages