计算机科学与探索 ›› 2025, Vol. 19 ›› Issue (1): 169-186.DOI: 10.3778/j.issn.1673-9418.2310047

• 理论·算法 • 上一篇    下一篇

面向异构社交网络的空-频域自适应图神经网络

张岚泽,顾益军,彭竞杰   

  1. 中国人民公安大学 信息网络安全学院,北京 100038
  • 出版日期:2025-01-01 发布日期:2024-12-31

Spatial-Frequency Domain Adaptive Graph Neural Network for Heterophilic Social Networks

ZHANG Lanze, GU Yijun, PENG Jingjie   

  1. School of Information and Cyber Security, People􀆳s Public Security University of China, Beijing 100038, China
  • Online:2025-01-01 Published:2024-12-31

摘要: 传统GNN基于同构性假设对近邻节点实现低通滤波功能完成邻域相似信息的聚合嵌入。但在异构图中分属不同类别的节点彼此多建立联系,而相同类别的节点在图拓扑位置上距离较远。这一特点给注重近端邻域信息聚合的传统GNN带来“远端节点信息聚合缺失”与“同构性假设失灵”的问题。因此设计融合空域与频域自适应嵌入机制的异构图神经网络(DA-HGNN)以解决上述问题。针对问题一,设计“远端空域嵌入模块”,旨在通过高阶随机游走迁移概率筛选并聚合远端相似节点,补充“消息传递的跨邻域自适应性”;针对问题二,设计“近端频域嵌入模块”,构建滤波器分离节点高频与低频信号,并设计频域导向型注意力机制对上述信息进行频域偏好的自适应融合,从而减少“同构性假设失灵”所引入的噪声。在四个公开异构图数据集中取得最优实验结果,准确率上平均提高6.41个百分点。在灵敏度分析和消融实验中阐述了超参数的选择机制和各模块的实际性能,并验证了在异构网络中“节点结构相似性”“节点属性向量相似性”以及“节点同构性”三者之间仍呈现正相关关系这一结论。在异构真实数据集中验证了欺诈检测效果,AUC指标提升4.4个百分点。

关键词: 异构图, 图神经网络, 图表示学习, 同构性假设失灵

Abstract: Traditional GNNs rely on the homophily assumption to implement low-pass filtering of neighboring nodes to aggregate and embed neighborhood similarity information. However, in heterophilic graphs, nodes belonging to different categories have many connections with each other, while nodes of the same category are far apart in the graph topology. This characteristic brings problems of “missing information aggregation of distant nodes” and “failure of homophily assumption” to traditional GNNs that focus on aggregating information in the proximal neighborhood. Therefore, this paper designs a heterophilic graph neural network (DA-HGNN) with a fusion of spatial-domain and frequency-domain adaptive embedding mechanisms to solve the above problems. To address the first problem, this paper designs a “distant spatial-domain embedding module” aimed at supplementing “cross-neighbor adaptive messaging” through high-order random walk migration probability selection and aggregation of distant similar nodes. To address the second problem, this paper develops a “proximal frequency-domain embedding module” to separate high-frequency and low-frequency signals using filters and designs a frequency-domain-guided attention mechanism to adaptively integrate the aforementioned information based on frequency preferences, thereby reducing the noise introduced by the “failure of homophily assumption”. The best experimental results are obtained on 4 publicly available heterophilic graph datasets, with an average increase in accuracy of 6.41 percentage points. Sensitivity analysis and ablation experiments describe the mechanism for selecting hyperparameters and the actual performance of each module, and verify the positive correlation among “node structural similarity” “node attribute vector similarity” and “node homophily” in heterophilic networks. Finally, the effectiveness of fraud detection is validated on a heterophilic real-world dataset, achieving an improvement of 4.4 percentage points in the AUC metric.

Key words: heterophilic graph, graph neural network, graph representation learning, failure of homophily assumption