Apprentissage de SVM sur Données Bruitées

Guillaume Stempfel (1) , Liva Ralaivola (1) , François Denis (1)

1 LIF - Laboratoire d'informatique Fondamentale de Marseille - UMR 6166

In this paper, we provide a simple example which shows that soft margin support vector machines (CSvm) are not tolerant to uniform classification noise. In order to cope with this limitation, we propose a noise-tolerant version of CSvm which is based on an objective function that makes use of an estimator of the noise-free slack margins. The nice properties of this estimator are supported by a theoretical analysis as well as numerical simulations carried out on a synthetic dataset.

Mots clés: Classification Supervisée; Séparateurs Linéaires; Machines à Vecteurs de Support; Bruit de Classification Uniforme

Domaines: Apprentissage [cs.LG]

Guillaume Stempfel, Liva Ralaivola, François Denis. Apprentissage de SVM sur Données Bruitées. 2008. ⟨hal-00256503⟩