Graph Domain Adaptation via Theory-Grounded Spectral Regularization.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Transfer Learning

Background:

• Transfer learning on graphs across different domains (GDA) is crucial for many applications.
• Current graph neural network (GNN) methods for GDA lack theoretical grounding and have variable performance.
• Existing approaches struggle to learn domain-invariant representations effectively.

Purpose of the Study:

• To develop theory-grounded algorithms for graph domain adaptation (GDA).
• To establish a theoretical foundation for improving GNN transferability across graph domains.
• To design GNNs capable of capturing more transferable representations.

Main Methods:

• Derived a model-based GDA bound using GNN spectral properties: spectral smoothness (SS) and maximum frequency response (MFR).
• Cross-pollinated optimal transport-based domain adaptation and graph filter theories.
• Proposed algorithms that regularize SS and MFR properties for enhanced GNN transferability.

Main Results:

• Developed a theoretical framework linking GNN spectral properties to GDA performance.
• Showcased that SS and MFR regularization benefit node and link transfer scenarios, respectively.
• Empirically validated the proposed methods on real-world datasets, demonstrating improved transferability.

Conclusions:

• The proposed spectral regularization techniques provide a theoretically sound approach to improve GNN transferability in domain adaptation.
• The study offers insights into tailoring regularization strategies for specific transfer learning tasks (node vs. link transfer).
• This work paves the way for constructing more robust and adaptable GNNs for cross-domain graph learning.