Rank and Align: Towards Effective Source-free Graph Domain Adaptation

TL;DR

The paper tackles source-free graph domain adaptation by transferring pre-trained GNNs to target graphs without accessing source data. It introduces Rank and Align (RNA), a framework that uses spectral seriation to derive robust target-semantic representations, detects harmonic graphs close to the source, and aligns inharmonic graphs through adversarial subgraph extraction and invariant learning, aided by filtered pseudo-labeling. The approach is supported by theoretical robustness guarantees for SSR under noise and validated through comprehensive experiments showing state-of-the-art performance and clear ablation-based gains from each component. The results suggest that combining spectral rankings, domain-invariant subgraph learning, and multi-view pseudo-label filtering yields practical, scalable improvements for real-world graph SFDA settings.

Abstract

Graph neural networks (GNNs) have achieved impressive performance in graph domain adaptation. However, extensive source graphs could be unavailable in real-world scenarios due to privacy and storage concerns. To this end, we investigate an underexplored yet practical problem of source-free graph domain adaptation, which transfers knowledge from source models instead of source graphs to a target domain. To solve this problem, we introduce a novel GNN-based approach called Rank and Align (RNA), which ranks graph similarities with spectral seriation for robust semantics learning, and aligns inharmonic graphs with harmonic graphs which close to the source domain for subgraph extraction. In particular, to overcome label scarcity, we employ the spectral seriation algorithm to infer the robust pairwise rankings, which can guide semantic learning using a similarity learning objective. To depict distribution shifts, we utilize spectral clustering and the silhouette coefficient to detect harmonic graphs, which the source model can easily classify. To reduce potential domain discrepancy, we extract domain-invariant subgraphs from inharmonic graphs by an adversarial edge sampling process, which guides the invariant learning of GNNs. Extensive experiments on several benchmark datasets demonstrate the effectiveness of our proposed RNA.

Figures (4)

• Figure 1: Motivation of RNA. (a) The histogram of confidence score of the source-trained model perform on source and target data, highlighting the label scarcity and potential noise in the target domain. (b) The training process of RNA and adaptation with pseudo-label learning. (c) Domain discrepancy between source and target domain. Best viewed in color and zoom-in.
• Figure 2: Overview of RNA. (Left) Rank. RNA use the seriation similarity ranking learning for robust semantics learning under label scarcity, as in Section \ref{['sec:SSR']}. (Right) Align. RNA detect the harmonic set and align the inharmonic set with a subgraph extractor and invariant learning, while applying discrimination learning with filtered pseudo-label, as in Section \ref{['sec:subgraph-align']} and \ref{['sec:pseudo-label']}. (Center) RNA alternates between two steps, achieving effective source-free domain adpation and addressing critical dilemmas.
• Figure 3: Visualization of RNA on Mutagenicity.
• Figure 4: Sensitivity analysis.

Theorems & Definitions (2)

• Theorem 3.1
• Theorem 3.2