TO-UGDA: target-oriented unsupervised graph domain adaptation.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Graph domain adaptation (GDA) faces challenges with limited labeled data in target graph domains.
• Existing GDA methods often rely solely on representation alignment, which can be affected by irrelevant information and ignore conditional shifts.

Purpose of the Study:

• To propose a target-oriented unsupervised graph domain adaptive framework (TO-UGDA) to effectively address limitations in GDA.
• To improve the transferability of label information from labeled source domains to unlabeled target domains.

Main Methods:

• Extracting domain-invariant feature representations using a graph information bottleneck.
• Minimizing domain discrepancy via adversarial alignment for a unified feature distribution.
• Employing meta pseudo-labeling to enhance downstream adaptation and model generalizability.

Main Results:

• The proposed TO-UGDA framework demonstrates excellent performance across various node-level and graph-level adaptation tasks.
• Experiments on real-world graph datasets validate the effectiveness of the framework.

Conclusions:

• TO-UGDA offers a robust solution for unsupervised graph domain adaptation.
• The framework effectively handles conditional shifts and irrelevant information, improving model generalizability.