A survey on graph structure learning

doi:10.1016/j.neucom.2026.134114

A survey on graph structure learning

2026-05-27

0

PRE

AI

DOI：10.1016/j.neucom.2026.134114

Original

Original request for help

Save

Abstract

En 中文

Graph Neural Networks (GNNs) have shown strong performance across diverse graph-based tasks, yet their accuracy can degrade substantially when the input graph is noisy, incomplete, or adversarially perturbed. Graph Structure Learning (GSL) addresses this problem by refining or inducing graph topology to better support representation learning and downstream objectives. Since 2023, the field has expanded through Graph Transformers, LLM-enhanced methods, and causal structure discovery, while the boundaries between paradigms have become less clear. Existing surveys cover only parts of this rapidly evolving landscape. This survey reviews GSL advances with primary coverage through early 2025. We examine GSL through four theoretical lenses spanning graph signal processing, biological plausibility, dynamical systems, and optimization theory, organize more than 80 methods into a two-dimensional framework with explicit classification criteria that clarify recurring boundary cases, and synthesize benchmark evidence across 19 datasets. Across unified evaluations, GSL is most useful on noisy, heterophilous, or incomplete graphs, while often yielding smaller gains on clean homophilous benchmarks. The survey concludes with practical guidance for method selection and a research roadmap.

Keywords:

Graph Structure Learning

Graph Neural Networks

Noisy Graphs

Heterophilous Graphs

Benchmark Evaluation

A survey on graph structure learning

Abstract

Journal

Organization

Cited Papers