Integration of Discriminate Features and Similarity Preserving for Unsupervised Domain Adaptation

Visual domain adaptations utilize source domain knowledge to train robust classifiers for the target domain. Reducing the gap between domain feature distributions could benefit from unsupervised domain adaptation. We have sufficient labeled source and unlabeled target domains in this scenario, but they come from distinct distributions while utilizing the same label space. As a result, applying information directly from the source to the target domain gives poor performance. To solve the issues, we propose the method IDFSP (Integrated Discriminative Features and Similarity Preserving) for unsupervised domain adaptation. In this paper, we apply entropy regularization for clustering and adapting the neighbors of unlabeled target domains to solve the problem. The discriminative and domain-invariant features for the source and target domains are combined. We consider retaining the discrimination and likeness of domains by viewing the linear and non-linearity data. We consider retaining the discrimination and likeness of domains by viewing the linear and non-linearity data. We carried out extensive experiments and confirmed our approach outperforms numerous state-of-the-art domain adaptation methods on four cross-domain visual tasks and the Amazon review sentiment analysis task.