Heterogeneous Graph Attention Network for Unsupervised Multiple-Target Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation typically focuses on single source-single target scenarios.
• Existing methods are suboptimal for multi-target domain adaptation due to ignoring inter-domain semantic associations.

Purpose of the Study:

• To propose a deep semantic information propagation approach for one-source, multiple-unlabeled-target domain adaptation.
• To learn a unified subspace common for all domains, enabling effective knowledge transfer.

Main Methods:

• Utilized a heterogeneous graph attention network (GAT) for semantic propagation across multiple domains.
• Employed an attention mechanism to optimize sample relationships for enhanced semantic transfer.
• Generated pseudo-labels for target domains to learn domain-invariant representations via centroid alignment.

Main Results:

• The proposed model successfully learns a unified subspace for all domains.
• Semantic information propagation among multiple target domains was achieved.
• Outperformed several popular domain adaptation methods on four public datasets.

Conclusions:

• The deep semantic information propagation approach is effective for one-source, multiple-target domain adaptation.
• Graph attention networks facilitate semantic transfer in complex domain adaptation settings.
• The method offers a promising solution for scenarios with limited labeled data across multiple domains.