Algorithm and Theoretical Analysis for Domain Adaptation Feature Learning with Linear Classifiers

Domain adaptation problem arises in a variety of applications where the training set (source domain) and testing set (target domain) follow different distributions. The difficulty of such learning problem lies in how to bridge the gap between the source distribution and target distribution. In this paper, we give an formal analysis of feature learning algorithms for domain adaptation with linear classifiers. Our analysis shows that in order to achieve good adaptation performance, the second moments of source domain distribution and target domain distribution should be similar. Based on such a result, a new linear feature learning algorithm for domain adaptation is designed and proposed. Furthermore, the new algorithm is extended to have multiple layers, resulting in becoming another linear feature learning algorithm. The newly introduced method is effective for the domain adaptation tasks on Amazon review dataset and spam dataset from ECML/PKDD 2006 discovery challenge.