Domain adaptation remains a significant challenge in artificial intelligence, especially when models trained in one domain are required to perform well in another. Conventional adversarial domain adaptation approaches typically focus on aligning individual samples but overlook collective interactions, resulting in reduced performance when dealing with noisy or ambiguous data. Addressing these limitations, Shikai Chen et al. from Southeast University and Lenovo Research published their novel findings on Collective Domain Adversarial Learning (ColDA) on 15 December 2025 in Frontiers of Computer Science, a journal co-published by Higher Education Press and Springer Nature.
The ColDA framework uniquely leverages collective interactions among groups of samples rather than individual instances. This set-level adversarial approach uses a newly proposed Dual-Set Domain Discriminator, equipped with attention mechanisms to simultaneously capture intra-domain and inter-domain relationships. By modeling these collective interactions, ColDA significantly improves the alignment of features across domains, enhancing the robustness and effectiveness of the domain adaptation process.
Experimental validation on widely recognized benchmark datasets, including VisDA-2017 and Office-Home, confirms the effectiveness of ColDA. Results demonstrate that ColDA consistently outperforms existing state-of-the-art methods, particularly in scenarios characterized by domain ambiguity and noisy data.
Future work includes expanding this approach to multi-domain and open-set domain adaptation, aiming to further enhance model robustness and generalization across even broader applications.
