Abstract
Unsupervised graph domain adaptation (UGDA) aims to transfer knowledge learned from labeled source graph datasets to unlabeled target graph datasets originating from a different distribution. Existing approaches often rely on spatial message-passing mechanisms, potentially overlooking the valuable information encoded in the spectral domain specific to UGDA tasks. This work begins with an empirical investigation revealing that low-frequency components of the graph spectrum tend to capture shared features across domains, whereas high-frequency components often represent domain-specific characteristics. Addressing the challenge of effectively utilizing these spectral insights, we introduce the Spectral Synergy Network (SSN), a novel framework for UGDA. SSN explicitly disentangles low- and high-frequency graph signals to capture both global structural similarities and local, detailed differences, thereby enriching graph-level representations. For the shared low frequency components, we employ a mutual information maximization objective across domains to encourage the learning of generalizable features. Concurrently, to address domain discrepancies highlighted by high-frequency signals, we implement a cross-domain contrastive learning strategy focusing on these components. By synergistically optimizing both low- and high-frequency information, SSN learns representations that are simultaneously discriminative for the task and invariant to the domain shift. Comprehensive experiments on various benchmark datasets validate the effectiveness of SSN, demonstrating significant improvements over existing state-of-the-art UGDA methods.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"Leveraging Spectral Graph Properties for Unsupervised Cross-Domain Graph Classification", Technical Disclosure Commons, (May 01, 2025)
https://www.tdcommons.org/dpubs_series/8064