Sélection de variables en apprentissage d’ordonnancement. Évaluation des SVM pondérés

Abstract : To select the most useful and the least redundant features to be used in ranking function to reduce computational costs is an issue in learning to rank (LTR). Regularized SVM are promising approaches in this context. In this paper, we propose new feature selection algorithms for LTR based on weighted SVM. We investigate an ℓ2-AROM algorithm to solve the ℓ0 norm problem and a weighted ℓ2 algorithm to solve ℓ0 et ℓ1 norm problems. Experiments on benchmarks and commercial datasets show that our algorithms are up to 10 times faster and use up to 7 times less features than state-of-the-art methods, with similar ranking performance.
Document type :
Journal articles
Complete list of metadatas

https://hal.inria.fr/hal-01259566
Contributor : Léa Laporte <>
Submitted on : Wednesday, January 20, 2016 - 4:01:52 PM
Last modification on : Friday, October 11, 2019 - 8:22:49 PM

Identifiers

Citation

Léa Laporte, Sébastien Déjean, Josiane Mothe. Sélection de variables en apprentissage d’ordonnancement. Évaluation des SVM pondérés. Revue des Sciences et Technologies de l'Information - Série Document Numérique, Lavoisier, 2015, 18 (1), pp.97-121. ⟨10.3166/dn.18.1.97-121⟩. ⟨hal-01259566⟩

Share

Metrics

Record views

477