Lecture

Domain Adaptation (DA) aims at learning a model from source data and at
using it to classify target data drawn according to a different statistical
distribution. In this work, we assume that the learning algorithm receives
labeled source examples and unlabeled target data. We present DABoost which
(i) takes its origin from both the theory of DA and the theory of boosting
and (ii) jointly minimizes the classification error over the source domain
and the proportion of margin violations over the target domain. We present
a  theoretical analysis about the convergence of our algorithm and derive
a bound on the target generalization error. We show practical evidences of
the efficiency of DABoost.