Rethinking structural semantics for unsupervised graph domain adaptation

doi:10.1016/j.knosys.2025.114878

Rethinking structural semantics for unsupervised graph domain adaptation

2025-11-13

0

PRE

AI

DOI：10.1016/j.knosys.2025.114878

Original

Original request for help

Save

Abstract

En 中文

Unsupervised graph domain adaptation (UGDA) is pivotal in the field of graph learning, which aims to transfer knowledge from a labeled source graph to an unlabeled target graph by reducing the distribution discrepancies. However, most existing approaches predominantly focus on learning domain-invariant node representations, neglecting structure shift caused by discrepancies of node connectivity between source and target graphs. This oversight often results in suboptimal knowledge transfer. To address this challenge, a novel method termed Rethinking Structural Semantics for Unsupervised Graph Domain Adaptation (S2GDA) is proposed, which explicitly mines and aligns diverse structural semantics for graph domain adaptation. Specifically, S2GDA first learns structural semantics by mapping the connection patterns to the multi-channel probability distribution between nodes, and then aggregates information through a semantic structure-guided message propagation mechanism to obtain robust node embeddings. Finally, node embeddings and structural semantics are jointly aligned to achieve effective knowledge transfer between graph domains. Theoretical analysis further demonstrates that explicit structural alignment enables a tighter adaptation bound. Extensive experiments on multiple real-world datasets show that S2GDA significantly outperforms the latest baseline models, with relative ranging from 0.33 % to 13.6 % across fourteen cross-graph node classification tasks, highlighting the significance of rethinking structural semantics for effective graph domain adaptation.

Rethinking structural semantics for unsupervised graph domain adaptation

Abstract

Journal

Organization

Cited Papers