Structured Domain Adaptation

In many real-world applications, labeled data are either expensive or too scarce to be used to train an accurate classifier. Therefore, it is worth exploring and often essential to make full use of existing resources. Domain adaptation is one of the most promising techniques of leveraging an existing well-labeled source domain and a limited labeled target domain. With the aim of better understanding new or unknown domains through well-labeled ones, this paper proposes an efficient and powerful algorithm, named structured domain adaptation (SDA), which transfers knowledge across two domains. Specifically, SDA aims to seek a discriminate subspace shared by two domains where the well-learned knowledge of the source domain can be transferred to the target domain. In SDA, samples from both domains are combined together to reveal more shared information across two domains. Furthermore, an iteratively structured matrix <tex-math notation="LaTeX">H </tex-math> is learned to bridge the domain shift so that the marginal and conditional distributions are mitigated. Since the target is labeled to a limited extent or even totally unlabeled, we adopt pseudolabels of the target data to optimize <tex-math notation="LaTeX">H </tex-math> iteratively so that a reconstruction coefficient matrix is learned in order to guarantee local awareness. Our approach is robust to outliers as it applies an <tex-math notation="LaTeX">2,1 </tex-math> -norm on the error term. SDA can work in either an unsupervised or a semisupervised manner. Extensive experiments on five data sets including faces, objects, digits, and visual events, demonstrate that SDA outperforms several state-of-the-art approaches with significant advantages. Notably, SDA can even achieve 100% accuracy on several popular benchmarks.