| 摘要: |
| 图自监督学习旨在无需人工标注的条件下学习有效的图结构表示. 尽管图对比学习(graph contrastive learning, GCL)通过构造标签保持不变的扰动视图并最大化其与原始视图的相似性来实现自监督训练, 但现有方法普遍采用全局均匀扰动策略, 未能考虑真实网络中节点角色的异质性——这种无差别的随机破坏会违反标签不变性假设. 实际观测表明, 真实网络普遍呈现核心-边缘的双层拓扑架构: 核心节点通过高度互连形成信息枢纽, 对其进行破坏性扰动将导致语义失真与标签偏移. 针对上述问题, 提出基于核心-边缘结构感知的图对比学习框架, 其创新性体现在: (i)摒弃全局均匀扰动范式, 通过核心-边缘检测算法精准定位边缘节点, 实施局部扰动以保持核心拓扑完整性, 严格遵循标签不变性原则; (ii)设计了边缘节点删除等增强操作, 模拟真实网络的动态演化过程, 强制模型捕获拓扑稳定性与噪声鲁棒性特征; (iii)构建了核心-边缘对比损失函数, 对不同结构重要性的节点在损失计算中赋予差异化权重, 有效引导模型关注核心信息并抑制边缘潜在的负面干扰. 在多个基准数据集上的实验表明, 所提方法在多个任务中均显著优于现有最优模型. |
| 关键词: 图对比学习 自监督学习 表示学习 |
| DOI：10.13328/j.cnki.jos.007579 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(62432003, 92267206) |
|
| Graph Contrastive Learning Method Based on Core-periphery Structure awareness |
|
WANG Ye-Jiang1,2, ZHAO Yu-Hai1,2, HUANG Miao-Miao1, WANG Mei-Xia1, LI Fang-Ting1, WANG Xing-Wei1
|
|
1.School of Computer Science and Engineering, Northeastern University, Shenyang 110167, China;2.Key Laboratory of Intelligent Computing of Medical Images (Northeastern University), Ministry of Education, Shenyang 110167, China
|
| Abstract: |
| Graph self-supervised learning aims to acquire effective graph structural representations without manual annotations. Although graph contrastive learning (GCL) enables self-supervised training by generating label-preserving perturbed views and maximizing their similarity to the original view, existing methods predominantly adopt a globally uniform perturbation strategy, which neglects the heterogeneity of node roles in real-world networks. Such indiscriminate random corruption violates the label-invariance assumption. Empirical observations indicate that real-world networks generally exhibit a core-periphery bi-layered topological architecture: core nodes form highly interconnected information hubs, and destructive perturbations applied to them are likely to cause semantic distortion and label shifts. To address these limitations, this study proposes a core-periphery structure-aware graph contrastive learning framework, with the following innovations: (i) instead of global uniform perturbation paradigm, peripheral nodes are accurately identified via a core-periphery detection algorithm, and localized perturbations are applied to preserve the integrity of the core topology, thereby strictly satisfying the principle of label invariance; (ii) augmentation operations such as peripheral node deletion are designed to simulate the dynamic evolution of real-world networks, encouraging the model to capture topological stability and noise robustness; (iii) a core-periphery contrastive loss function is constructed, in which differentiated weights are assigned to nodes with varying structural importance during loss computation, effectively guiding the model to emphasize core information while suppressing potential negative interference from peripheral nodes. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method consistently outperforms state-of-the-art models across various tasks. |
| Key words: graph contrastive learning (GCL) self-supervised learning representation learning |
