Domain Adaptation with Few Labeled Source Samples by Graph Regularization

Domain Adaptation aims at utilizing source data to establish an exact model for a related but different target domain. In recent years, many effective models have been proposed to propagate label information across domains. However, these models rely on large-scale labeled data in source domain and cannot handle the case where the source domain lacks label information. In this paper, we put forward a Graph Regularized Domain Adaptation (GDA) to tackle this problem. Specifically, the proposed GDA integrates graph regularization with maximum mean discrepancy (MMD). Hence GDA enables sufficient unlabeled source data to facilitate knowledge transfer by utilizing the geometric property of source domain, simultaneously, due to the embedding of MMD, GDA can reduce source and target distribution divergency to learn a generalized classifier. Experimental results validate that our GDA outperforms the traditional algorithms when there are few labeled source samples.