On learning a large margin classifier for domain adaptation based on similarity functions
Résumé
Traditional supervised classification algorithms fail when unlabeled test data arise from a probability distribution that differs from that of the labeled training data. This problem is addressed by domain adaptation , an active research area in which one would like to transfer the knowledge acquired from a first labeled domain, the source, to a second one, the target. In this paper, we tackle this problem from the perspective of large margin classifiers based on (, γ, τ)−good similarity functions. We first prove a bound on the error of such a classifier on the target domain. Then we present our algorithm consisting in minimizing this bound, allowing to learn a good classifier directly on the target domain without an intermediate domain alignment step. Under specific conditions, our algorithm can be formulated as a convex optimization problem that is solved efficiently. Its performance is assessed via experiments on on a toy set and a real world problem.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...