GALA: Graph Diffusion-based Alignment with Jigsaw for Source-free Domain Adaptation.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Domain Adaptation

Background:

• Source-free domain adaptation is vital for data privacy, but existing methods neglect non-Euclidean graph data.
• Graph Neural Networks (GNNs) face performance issues in source-free scenarios due to domain shift and limited labels.

Purpose of the Study:

• To propose GALA (Graph Diffusion-based Alignment with Jigsaw), a novel method for source-free graph domain adaptation.
• To address the performance decline of GNNs in scenarios with domain shift and label scarcity.

Main Methods:

• GALA employs a graph diffusion model trained on source graphs to reconstruct source-style graphs from target data.
• A stochastic differential equation introduces perturbations for graph reconstruction, followed by pseudo-label generation using class-specific thresholds and curriculum learning.
• A graph jigsaw strategy enhances generalization and robustness through consistency learning by mixing confident and unconfident graphs.

Main Results:

• GALA effectively aligns domains by reconstructing source-style graphs from target data.
• The method generates accurate and unbiased pseudo-labels for target graphs.
• The graph jigsaw strategy improves generalization and robustness.

Conclusions:

• GALA demonstrates significant effectiveness in source-free graph domain adaptation.
• The proposed method offers a robust solution for handling domain shift and label scarcity in graph data.
• GALA advances the application of machine learning in privacy-preserving, real-world scenarios involving graph data.