Unsupervised domain adaptation with post-adaptation labeled domain performance preservation

Unsupervised domain adaptation is a machine learning-oriented application that aims to transfer knowledge learned from a seen (source) domain with labeled data to an unseen (target) domain with only unlabeled data. Recently developed techniques apply adversarial learning to learn domain-transferable features. However, current adversarial domain adaptation models suffer from the training instability of adversarial networks. Furthermore, it is unclear what the source domain pays in terms of performance during learning the domain-transferable representation. To address this issue, we propose a novel approach termed U nsupervised D omain A daptation with S ource P reservation (UDA-SP). It shares the same objective of obtaining a generalization representation between different distributions as domain adaptation techniques. Additionally, it has the new objective of preserving efficient performance in the source domain. This is accomplished by learning representations of shared and source-specific features that are separately learned from two distinct networks. Then, they are concatenated with available class information to train a new classifier that has the ability to exploit both shared and domain-specific features. We conducted a comprehensive experimental analysis on three benchmark text datasets. Experiments validate that our proposed method outperforms their competing state-of-the-art methods. Further experiments demonstrate that UDA-SP has a good ability to generalize learned knowledge to unseen domains while maintaining seen domain performance.