[论文解读] A Survey on Graph Structure Learning: Progress and Opportunities

Yanqiao Zhu, Weizhi Xu+6 more|arXiv (Cornell University)|Mar 4, 2021

Advanced Graph Neural Networks被引用 67

一句话总结

This survey overviews Graph Structure Learning (GSL), categorizes methods, outlines a general pipeline, and discusses applications and future challenges.

ABSTRACT

Graphs are widely used to describe real-world objects and their interactions. Graph Neural Networks (GNNs) as a de facto model for analyzing graphstructured data, are highly sensitive to the quality of the given graph structures. Therefore, noisy or incomplete graphs often lead to unsatisfactory representations and prevent us from fully understanding the mechanism underlying the system. In pursuit of an optimal graph structure for downstream tasks, recent studies have sparked an effort around the central theme of Graph Structure Learning (GSL), which aims to jointly learn an optimized graph structure and corresponding graph representations. In the presented survey, we broadly review recent progress in GSL methods. Specifically, we first formulate a general pipeline of GSL and review state-of-the-art methods classified by the way of modeling graph structures, followed by applications of GSL across domains. Finally, we point out some issues in current studies and discuss future directions.

研究动机与目标

• Formulate the Graph Structure Learning (GSL) problem and its significance for improving GNN performance on noisy or incomplete graphs.
• Present a general three-stage pipeline for GSL: graph construction, structure modeling, and message propagation.
• Classify GSL methods into metric-based, neural, and direct approaches with representative techniques.
• Discuss common graph regularizers and postprocessing steps, and survey diverse applications across domains.
• Outline current challenges and future research directions to guide further development in GSL.

提出的方法

• Define the GSL objective as jointly learning an optimized adjacency matrix and node representations.
• Propose a three-stage GSL pipeline: graph construction, graph structure modeling, and message propagation.
• Categorize structure modeling into metric-based, neural, and direct approaches with examples.
• Describe graph regularizers (sparsity, smoothness, community preservation) and postprocessing (discrete sampling, residual connections).
• Discuss training frameworks and strategies such as alternating optimization and bilevel/meta-learning formulations.
• Summarize cross-domain applications and provide directions for future research.

实验结果

研究问题

• RQ1 What is Graph Structure Learning and why is it needed for GNNs?
• RQ2 How can graph structures be constructed and refined to improve downstream tasks?
• RQ3 What are the main categories and representative methods of GSL, and how do they differ?
• RQ4 What are common regularizers and postprocessing techniques used in GSL?
• RQ5 What are the current challenges and future directions in GSL research?

主要发现

• GSL is framed as jointly learning a graph structure and its representations to enhance downstream performance.
• A three-stage pipeline (construction, modeling, propagation) is central to most GSL methods.
• Methods are categorized into metric-based, neural, and direct approaches, each with distinct pros and mechanisms.
• Graph regularizers such as sparsity, smoothness, and community preservation guide learned structures; postprocessing includes discrete sampling and residual connections.
• GSL has broad applicability across NLP, computer vision, medical imaging, and scientific discovery, with several open challenges and promising directions.