Cross-Domain Graph Convolutions for Adversarial Unsupervised Domain Adaptation

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) aims to adapt models to new data domains without labels.
• Adversarial methods in UDA create domain-invariant features but can distort data structures.
• Existing methods struggle with preserving both sample and class-level information during adaptation.

Purpose of the Study:

• To propose a novel transferable feature learning approach on graphs (TFLG) for unsupervised adversarial domain adaptation.
• To jointly align sample- and class-level structures across source and target domains.
• To enhance classifier adaptation in the presence of domain shift.

Main Methods:

• TFLG constructs graphs for minibatch samples to capture relationships.
• Identifies classwise correspondence between source and target domains.
• Employs a novel cross-domain graph convolutional operation and a memory bank for structure alignment.

Main Results:

• The proposed TFLG method demonstrates significant effectiveness in unsupervised domain adaptation.
• Achieves superior performance compared to state-of-the-art UDA techniques on benchmark datasets.
• Successfully preserves and aligns both sample- and class-level data structures.

Conclusions:

• TFLG offers an effective solution for unsupervised adversarial domain adaptation by integrating graph structures.
• Jointly aligning sample and class information improves feature transferability.
• The approach shows promise for real-world applications requiring adaptation to new data domains.