Unsupervised domain adaptation with non-stochastic missing data

We consider unsupervised domain adaptation (UDA) for classification problems in the presence of missing data unlabelled target domain. More precisely, motivated by practical applications, we analyze situations where distribution shift exists between domains and some components are systematically absent on without available supervision imputing components. propose a generative approach imputation. Imputation is performed domain-invariant latent space leverages indirect from complete source introduce single model performing joint adaptation, imputation which, under our assumptions, minimizes an upper bound its generalization error performs well various representative divergence families ( $$\mathscr {H}$$ -divergence, Optimal Transport). Moreover, compare adaptation-imputation framework “ideal” UDA classifier Our further improved with self-training, to bring learned class posterior distributions closer. perform experiments three datasets different modalities: classical digit benchmark, Amazon product reviews dataset both commonly used real-world digital advertising datasets. show benefits jointly these