基于双注意力图神经网络的链路预测  

Link prediction based on a dual attention graph neural network

作　　者：杨真真[1] 林泽龙 杨永鹏[1] YANG Zhenzhen;LIN Zelong;YANG Yongpeng(Key Laboratory of Broadband Wireless Communication and Sensor Network Technology,Ministry of Education,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学宽带无线通信与传感网技术教育部重点实验室,江苏南京210003

出　　处：《南京邮电大学学报(自然科学版)》2025年第1期106-114,共9页Journal of Nanjing University of Posts and Telecommunications:Natural Science Edition

基　　金：国家自然科学基金(62071242,62171232);江苏省研究生科研与实践创新计划(KYCX22_0955,SJCX23_0251)资助项目。

摘　　要：链路预测是在图结构中预测未知或潜在的边,对挖掘图中的隐含信息、补全图中的缺失数据和发现图中的新知识都具有重要意义。图神经网络(Graph Neural Network,GNN)已被广泛应用于链路预测,然而,现有基于GNN的链路预测方法存在一些问题:(1)大多数基于GNN的方法往往容易忽略为链路预测提供额外帮助的边信息的重要性;(2)大多数基于GNN的方法都仅捕获表示图的邻居节点间相似性的低频信息,忽略了表示邻居节点间差异性的高频信息;(3)大多数基于GNN的方法都未考虑输入特征矩阵的节点维度和特征维度两个维度,只关注其中一个维度。针对这些问题,提出了一种基于双注意力图神经网络(Dual Attention Graph Neural Network,DAGNN)的链路预测方法,该方法包含两条路径,以不同的角度更新节点表示。其中一条是基于图神经网络的路径,采用含边信息的频率自适应图注意力网络(Frequency Adaptive Graph Attention Network with Edge Information,FAGAT⁃EI)作为基础模型,有效地利用边信息增强节点之间的关系,并利用频率自适应机制平衡高低频率邻居信息的权重,从而缓解GNN的过度平滑问题;另一条是基于通道注意力网络的路径,提出了一种新的压缩-激励通道注意力模块(Squeeze and Excitation⁃Channel At⁃tention Module,SE⁃CAM)作为基础模型,充分考虑输入特征矩阵的节点维度和特征维度,并自动学习和调整每个节点的不同特征权重,从而得到更有意义的节点表示。最后在两个基准数据集上进行了实验,实验结果表明,提出的链路预测方法在Last⁃FM和Book⁃Crossing两个数据集上的AUC和ACC指标均优于其他基线模型,展现出了卓越的链路预测性能。Link prediction is the process of predicting unknown or potential edges in a graph structure,which is of great significance for mining hidden information,completing missing data,and discovering new knowledge in the graph.Graph neural networks(GNNs)have been commonly used in link predic⁃tion.However,existing GNN-based link prediction methods have some problems:(1)Most of them often tend to ignore the importance of edge information that provides additional assistance for link prediction;(2)most of themonly capture the low-frequency information of the graph representing local similarity between neighboring nodes,but ignore the high-frequency information of the graph representing global dif⁃ferences between neighboring nodes;(3)most of them overlook both node and feature dimensions of the input feature matrix,focusing on its one dimension.To address these problems,we propose a dual atten⁃tion graph neural network(DAGNN)for link prediction.The proposed network comprises two paths to up⁃date node representations from different perspectives.One path is based on GNN.It utilizes a frequency adaptive graph attention network with edge information(FAGAT-EI)as the baseline,effectively incorpo⁃rating edge information to enhance relationships between nodes.The frequency-adaptive mechanism bal⁃ances the weights of low and high frequency neighbor information,alleviating the issue of oversmoothing.The other path is based on a channel attention network.A novel squeeze and excitationchannel attention module(SE-CAM)is introduced as the baseline.This path fully considers the node and feature dimensions of the input feature matrix,automatically learning and adjusting weights for different features of each node,resulting in more meaningful node representations.Finally,we conduct extensive experiments on two benchmark datasets.The experimental results verify the effectiveness and superiority of the proposed link prediction method,and demonstrate that it outperforms other methods in terms of AUC and ACC metrics on the Last-FM

关 键 词：链路预测 图神经网络 注意力机制 压缩-激励模块 频率自适应 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]