Learning domain-shared group-sparse representation for unsupervised domain adaptation

In unsupervised domain adaptation, a key research problem is joint distribution alignment across the source and target domains. However, direct alignment of the source and target joint distributions is infeasible, because the target conditional distribution cannot be known without target labels. Instead of estimating target labels for target conditional distribution approximation, this paper proposes a new criterion of domain-shared group sparsity that is an equivalent condition for conditional distribution alignment. Together with marginal distribution alignment, we develop a domain-shared group-sparse dictionary learning model to learn domain-shared representations with aligned joint distributions. A cross-domain label propagation method is then proposed to train a classifier for the target domain using the domain-shared group-sparse representations and the target-specific information from the target data. The proposed method outperforms eight state-of-the-art unsupervised domain adaptation algorithms for cross-domain face recognition and cross-dataset object recognition with hand-drafted and deep features. Experimental results across multiple sub-domains show that the proposed method also performs well across datasets with large variance. Our results are quantitatively and qualitatively analyzed, and experiments of parameter sensitivity and convergence analysis are performed to show the effectiveness of the proposed method.