Abstract
Domain adaptation has been a fundamental technology for transferringknowledge from a source domain to a target domain. The key issue of domainadaptation is how to reduce the distribution discrepancy between two domains ina proper way such that they can be treated indifferently for learning.Different from existing methods that make label prediction for target samplesindependently, in this paper, we propose a novel domain adaptation approachthat assigns pseudo-labels to target data with the guidance of class centroidsin two domains, so that the data distribution structure of both source andtarget domains can be emphasized. Besides, to explore the structure informationof target data more thoroughly, we further introduce a local connectivityself-learning strategy into our proposal to adaptively capture the inherentlocal manifold structure of target samples. The aforementioned class centroidmatching and local manifold self-learning are integrated into one jointoptimization problem and an iterative optimization algorithm is designed tosolve it with theoretical convergence guarantee. In addition to unsuperviseddomain adaptation, we further extend our method to the semi-supervised scenarioincluding both homogeneous and heterogeneous settings in a direct but elegantway. Extensive experiments on five benchmark datasets validate the significantsuperiority of our proposal in both unsupervised and semi-supervised manners.