Sparsely-labeled source assisted domain adaptation

Domain Adaptation (DA) aims to generalize the classifier learned from a well-labeled source domain an unlabeled target domain. Existing DA methods usually assume that rich labels could be available in However, we confront with large number of data but only few labeled data, and thus, how transfer knowledge this sparsely-labeled is still challenge, which greatly limits its application wild. This paper proposes novel Sparsely-Labeled Source Assisted (SLSA-DA) algorithm address challenge limited samples. Specifically, due label scarcity problem, projected clustering first conducted on both domains, so discriminative structures exploited elegantly. Then propagation adopted propagate those samples whole progressively, cluster are revealed correctly. Finally, jointly align marginal conditional distributions mitigate cross-domain mismatching optimize three procedures iteratively. it nontrivial incorporate above into unified optimization framework seamlessly since some variables optimized implicitly involved their formulas, thus they not benefit each other. Remarkably, prove distribution alignment reformulated other formulations, implicit embedded different steps. As such, related quantities other, improve recognition performance obviously. Extensive experiments have verified our approach deal SLSA-DA setting, achieve best performances across real-world visual tasks. Our preliminary Matlab code at https://github.com/WWLoveTransfer/SLSA-DA/.