Unsupervised domain adaptation via feature transfer learning based on elastic embedding

Supervised classification algorithms usually require a large quantity of well-labeled samples for training to achieve satisfied performance. Nevertheless, it is prohibitively difficult to create such datasets with complete annotation. Unsupervised domain adaptation is able to deal with this problem via transferring knowledge from a relevant dataset with rich labels to an unlabeled target dataset. In this paper, a novel unsupervised domain adaptation method named feature transfer learning based on elastic embedding (EEFTL) is presented for image classification. Rather than make a rigid embedding that the binary label matrix is exactly equal to a linear function, EEFTL adopts an elastic embedding induced by the prediction label matrix. Specifically, EEFTL introduces a flexible regression residue term to model the mismatch between the embedded features of samples and the prediction labels. In addition, EEFTL integrates a label fitness term to effectively utilize the label information from the source samples, a distribution matching term to reduce the distances between domains in both the marginal and conditional distributions, and a manifold regularization term to preserve the sample-wise structure information under the elastic embedding. Extensive experiments are carried out on multiple benchmark datasets, and the results prove the effectiveness of the proposed method.