Multi-source domain adaptation-based low-rank representation and correlation alignment

Domain adaptation is one of the machine learning approaches, which is very powerful and applicable especially when there is no labeled data on the target domain or there are unequal distributions and different feature spaces between the source and target domains. This paper proposes an unsupervised domain adaptation model, which addresses this problem by utilizing two main folds: first, domain shift between source and target is diminished by matching the second-order statistics of distributions, and then the aligned source data along with target data are transferred into a shared subspace where more reduction in distributions discrepancy is occurred by linear combinations of related source samples to each target sample with utilizing low-rank and sparse conditions. So the classification ability of the source domain is transferred into the target domain. The experimental results mention that the proposed approach outperforms competitors.