Active domain adaptation seeks to cut annotation costs while improving unsupervised domain adaptation, but models optimized for average accuracy often neglect difficult categories. The authors argue that point-wise selection strategies built on influence functions can be sensitive and redundant, causing the acquisition function to overfit.

To address this, they introduce a region-based approach that maximizes positive manifold coverage, ensuring more robust improvement of the adaptation model. An uncertainty score is added to the selection mechanism to help learn a better decision boundary for the target domain.

In experiments across multiple domain adaptation benchmarks, the proposed method surpasses the first category-aware active domain adaptation baseline and existing active domain adaptation methods. The paper has been accepted at the Machine Learning Journal through the ACML Journal Track 2026, and the implementation is publicly available.