HG-SFDA: HyperGraph Learning Meets Source-Free Unsupervised Domain Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Source-Free unsupervised Domain Adaptation (SFDA) is crucial for classifying target data without source data access.
• Existing SFDA methods struggle with limited pairwise relations and overlook domain shift effects.

Purpose of the Study:

• To develop a new SFDA method that leverages high-order neighborhood relations and explicitly accounts for domain shift.
• To improve knowledge transfer from source to target domains in unsupervised settings.

Main Methods:

• Formulated SFDA as a hypergraph learning problem, constructing hyperedges to capture multi-sample structural information.
• Integrated a self-loop strategy to model sample domain uncertainty.
• Employed clustering based on hyperedges considering both semantic features and domain shift.
• Utilized an adaptive relation-based objective with soft attention for model tuning.

Main Results:

• The proposed hypergraph learning method demonstrated superior performance across multiple benchmark datasets (Office-31, Office-Home, VisDA, DomainNet-126, PointDA-10).
• The approach effectively addressed limitations of pairwise relation methods and semantic-feature-only clustering.
• Significant improvements were observed compared to state-of-the-art SFDA techniques.

Conclusions:

• The novel hypergraph learning approach offers a robust solution for SFDA by integrating structural information and domain shift awareness.
• This method enhances the accuracy and reliability of domain adaptation without source data.
• The findings pave the way for more effective unsupervised domain adaptation strategies.