In visual domain adaptation, the goal is to train effective classifiers for the target domain by leveraging information from the source domain. In unsupervised domain adaptation, the source domain provides labeled data while the target domain lacks labels. However, it is crucial to recognize that the source and target domains have different underlying distributions despite sharing the same label space. Directly applying source domain information to the target domain often leads to poor performance due to the distribution gap between the two domains. Unsupervised do- main adaptation aims to bridge this gap and improve performance. We introduce a comprehensive UDADFSP (Unified Domain Adaptation with Discriminative Features and Similarity Preservation) de- signed explicitly for unsupervised domain adaptation to tackle these challenges. Our framework focuses on incorporating discriminative and invariant features. We employ clustering with entropy regularization on the unlabeled target domain to refine the neighbor relationships. This step significantly enhances the alignment between the target and source domains, facilitating a more effective adaptation. Furthermore, we seamlessly incorporate discriminative features while preserving similarity in the source and target domains. We carefully balance the discrimination and similarity aspects by considering linear and non-linear data representations. Extensive testing demonstrates that learning discriminative and similarity features in the same feature space yields significant improvements over several state-of-the-art domain adaptation techniques. In a comparative evaluation, our approach surpasses several existing methods across four diverse cross-domain visual tasks and the Amazon re- view sentiment analysis task.