融合拉普拉斯位置编码和自注意力机制的图神经网络  

Graph neural network incorporating Laplacian position encoding and self-attention mechanism

作　　者：邹成龙 李伟诺 黄梅香 林艺东 ZOU Chenglong;LI Weinuo;HUANG Meixiang;LIN Yidong(School of Mathematics and Statistics,Minnan Normal University,Zhangzhou,Fujian 363000,China)

机构地区：[1]闽南师范大学数学与统计学院,福建漳州363000

出　　处：《福州大学学报(自然科学版)》2025年第2期127-134,共8页Journal of Fuzhou University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(12201284);福建省自然科学基金资助项目(2022J05169);闽南师范大学基金资助项目(KJ2021020&MSGJB2022010)。

摘　　要：为解决现有位置编码捕获图结构信息不足和谱图滤波器表达能力有限的问题,提出一种基于拉普拉斯位置编码和自注意力机制的LESpecformer模型架构.首先,引入拉普拉斯位置编码,用于感知不同节点在图结构中的位置信息,提升模型学习不同节点相对位置信息的能力,进而捕获图结构的全局信息.其次,基于融合拉普拉斯位置编码的特征值集合,利用自注意力机制自适应地学习特征值之间的依赖关系并获得有效的新基表示,使所提出的模型学习到更好的节点嵌入,从而提升节点分类精度.最后,在6个图数据集上,与不同基线网络的节点分类实验进行对比,结果表明所提出的LESpecformer性能均达到最优.In order to solve the problem of insufficient structure information captured by existing position encoding and limited expression ability of spectral graph filters,the LESpecformer model architecture based on Laplacian position encoding and self-attention mechanism is proposed.Firstly,this paper introduces the Laplacian position encoding for sensing the position information of different nodes in the graph structure,which improves the ability of the model to learn the relative position information of different nodes,and then captures the global information of the graph structure.Secondly,based on the set of eigenvalues fused with Laplacian position encoding,the self-attention mechanism is used to adaptively learn the dependencies between eigenvalues and obtain effective new basis representations,which facilitates the proposed model in learning better node embeddings,and thus improves the accuracy of node classification.Finally,compared with different baseline networks on six graph datasets,the experimental results show that the performance of the proposed LESpecformer is optimal.

关 键 词：谱图滤波器 拉普拉斯位置编码 自注意力机制 图神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]