A New PAC-Bayesian Perspective on Domain Adaptation

Pascal Germain (1) , Amaury Habrard (2) , François Laviolette (1) , Emilie Morvant (2)

1 Department of Computer Science and Software Engineering [Québec]
2 LHC - Laboratoire Hubert Curien

We study the issue of domain adaptation: we want to adapt a model from a source distribution to a target one. We focus on models expressed as a majority vote. Our main contribution is a novel theoretical analysis of the target risk that is formulated as an upper bound expressing a trade-off between only two terms: (i) the voters' joint errors on the source distribution, and (ii) the voters' disagreement on the target one; both easily estimable from samples. Hence, this new study is more precise than other analyses that usually rely on three terms (including a hardly controllable term). Moreover, we derive a PAC-Bayesian generalization bound, and specialize the result to linear classifiers to propose a learning algorithm.

Domaines: Machine Learning [stat.ML]

Pascal Germain, Amaury Habrard, François Laviolette, Emilie Morvant. A New PAC-Bayesian Perspective on Domain Adaptation. [Research Report] Laboratoire Hubert Curien, Université Jean Monnet, Saint-Etienne; Département d'informatique et de génie logiciel, Université Laval (Québec). 2015. ⟨hal-01163722v1⟩