Target-Adaptive Structure-Semantic Consistency for Unsupervised Graph Domain Adaptation

Abstract

Unsupervised Graph Domain Adaptation (UGDA) aims to mitigate distribution shifts between domains by transferring knowledge from labeled source graphs to unlabeled target graphs. Current work indicates that enhancing target embeddings is helpful for domain generalization. However, these methods primarily focus on structure-guided enhancement but often overlook the intrinsic coupling between structural topology and node semantics in graph data, resulting in suboptimal target representations during complex structure adaptation. To address this problem, we propose a novel approach called Target-adaptive Structure-Semantic Consistency (TASSC). First, we establish bidirectional optimization, ensuring consistency between structural proximity and semantic similarity on the target graph. Specifically, we propose a hybrid contrastive learning strategy, which unifies topological neighbors and cosine-similarity features (semantic neighbors) as positive samples. Additionally, we employ entropy minimization to suppress target semantic ambiguity caused by source domain biases, creating a closed-loop optimization where ‘structure guides semantics, semantics feedback structure.’ Furthermore, we develop a scale-aware adaptive module to access scale disparities between domains, dynamically transferring source knowledge to mitigate target semantic insufficiency. Extensive experiments on three real-world benchmark datasets demonstrate that our method achieves state-of-the-art results.