Abstract
Domain adaptation aims to transfer knowledge from a well-annotated source domain to a target domain with limited or no labeled data. A major challenge in this process is the domain shift problem, where the data distributions of the source and target domains differ significantly. Traditional domain adaptation methods primarily focus on re-weighting samples to align the data distributions, often neglecting the semantic relationships. Furthermore, previous approaches lack interpretability and fail to consider the performance of individual categories. In this dissertation, we propose novel frameworks that address these challenges by utilizing semantic alignment, visualizing transferred knowledge, and integrating human-in-the-loop strategies to improve domain adaptation performance. By integrating explainable AI and data-centric AI techniques, we offer new perspectives on domain adaptation, with potential applications including model diagnostics, large language model fine-tuning, and beyond.