Séparateurs à Vaste Marge pondérés en norme L2 pour la sélection de variables en apprentissage d'ordonnancement

Learning to rank algorithms are dealing with a very large amount of features to automatically learn ranking functions, which leads to an increase of both the computational cost and the number of noisy redundant features. Feature selection is seen as a promising way to address these issues. In this paper, we propose new feature selection algorithms for learning to rank based on reweighted `2 SVM approaches. We investigate a `2-AROM algorithm to solve the `0 norm optimization problem and a generic `2-reweighted algorithm to approximate `0 et `1 norm SVM problems with `2 norm SVM. Experiments show that our algorithms are up to 10 times faster and use up to 7 times less features than state-of-the-art methods, without lowering the ranking performance. MOTS-CLÉS : Apprentissage d’ordonnancement, Sélection de variables, SVM parcimonieux, Algorithmes de pondération.