[论文解读] Towards Effective and Efficient Graph Alignment without Supervision

Songyang Chen, Youfang Lin+3 more|arXiv (Cornell University)|Mar 9, 2026

Advanced Graph Neural Networks被引用 0

一句话总结

本文提出 GlobAlign，一种全局表示与传输为核心的无监督图对齐框架；以及通过对传输成本进行稀疏化的更高效变体 GlobAlign-E，在准确性和运行速度上均优于基于 OT 的方法。

ABSTRACT

Unsupervised graph alignment aims to find the node correspondence across different graphs without any anchor node pairs. Despite the recent efforts utilizing deep learning-based techniques, such as the embedding and optimal transport (OT)-based approaches, we observe their limitations in terms of model accuracy-efficiency tradeoff. By focusing on the exploitation of local and global graph information, we formalize them as the ``local representation, global alignment'' paradigm, and present a new ``global representation and alignment'' paradigm to resolve the mismatch between the two phases in the alignment process. We then propose \underline{Gl}obal representation and \underline{o}ptimal transport-\underline{b}ased \underline{Align}ment ( exttt{GlobAlign}), and its variant, exttt{GlobAlign-E}, for better \underline{E}fficiency. Our methods are equipped with the global attention mechanism and a hierarchical cross-graph transport cost, able to capture long-range and implicit node dependencies beyond the local graph structure. Furthermore, exttt{GlobAlign-E} successfully closes the time complexity gap between representative embedding and OT-based methods, reducing OT's cubic complexity to quadratic terms. Through extensive experiments, our methods demonstrate superior performance, with up to a 20\% accuracy improvement over the best competitor. Meanwhile, exttt{GlobAlign-E} achieves the best efficiency, with an order of magnitude speedup against existing OT-based methods.

研究动机与目标

• Formalize the limitations of the traditional two-phase "local representation, global alignment" paradigm in unsupervised graph alignment.
• Propose a new "global representation and alignment" paradigm to capture long-range and implicit node dependencies.
• Develop GlobAlign to leverage global information via self-attention and a hierarchical cross-graph transport cost.
• Introduce GlobAlign-E to reduce the time complexity gap between embedding-based and OT-based methods while preserving accuracy.
• Demonstrate superior accuracy and efficiency through extensive experiments and ablations.

提出的方法

• Formalize existing alignment methods into two phases: local representation and global alignment.
• Introduce global representations using self-attention (Transformers) to encode global graph information.
• Design a hierarchical cross-graph transport cost combining Gromov-Wasserstein and Wasserstein components.
• Optimize with an OT-based objective using alternating updates for the cost and the alignment matrix T, via proximal alternating linearized minimization and Sinkhorn iterations.
• Provide a sparsified variant GlobAlign-E to reduce cubic OT complexity to near-quadratic by masking relation matrices with top-k structure and semantic similarities.
• Analyze complexity and show GlobAlign-E achieves O(n^3) and GlobAlign-E achieves O(n^2 d + n m) under sparsification.

实验结果

研究问题

• RQ1 Can the mismatch between local representations and global alignment be resolved by adopting a global representation and transport paradigm?
• RQ2 How can global interactions be incorporated into unsupervised graph alignment to capture long-range dependencies?
• RQ3 Does a hierarchical cross-graph transport cost combining GW and WD improve accuracy and efficiency over existing OT-based methods?
• RQ4 Can a sparsified transport cost (GlobAlign-E) close the efficiency gap between embedding-based and OT-based methods without sacrificing performance?

主要发现

• GlobAlign achieves superior accuracy, with up to a 20% improvement over the best competitor.
• GlobAlign-E achieves best efficiency, with an order of magnitude speedup against existing OT-based methods.
• GlobAlign leverages global information via self-attention to model long-range dependencies beyond local graph structure.
• The hierarchical transport cost combines GWD and WD to balance global structure awareness and node-wise similarity.
• Sparsification via top-k masks and PPR-based structure, plus semantic masks, yields substantial efficiency gains with preserved accuracy.
• Empirical results include robustness analysis and ablation studies validating the effectiveness of the proposed paradigm and models.