一种近似图神经网络框架的无监督链路预测算法  

An Unsupervised Link Prediction Algorithm Based on an Approximate Graph Neural Network Framework

作　　者：李格格 冶忠林 曹淑娟 周琳 王雪力 LI Gege;YE Zhonglin;CAO Shujuan;ZHOU Lin;WANG Xueli(College of Computer,Qinghai Normal University,Xining 810008,China;The State Key Laboratory of Tibetan Intelligent Information Processing and Application,Qinghai Normal University,Xining 810008,China)

机构地区：[1]青海师范大学计算机学院,青海西宁810008 [2]青海师范大学藏语智能信息处理及应用国家重点实验室,青海西宁810008

出　　处：《郑州大学学报（工学版）》2024年第6期75-82,共8页Journal of Zhengzhou University（Engineering Science）

基　　金：国家重点研发计划(2020YFC1523300);青海省创新平台建设项目(2022-ZJ-T02)。

摘　　要：对于无标签网络,由于基于图神经网络的链路预测方法使用其高效建模机制进行链路预测任务时性能较差,因此,提出了一种近似图神经网络框架的无监督链路预测算法(ALIP),旨在模拟图神经网络算法的高效建模机制和学习过程,解决网络节点标签缺失导致的建模不充分问题。首先,参照GCN的输入层,融合网络的结构信息和节点属性;其次,使用矩阵分解替代GCN的隐藏层,模拟正向传播;再次,借鉴恒等映射和高阶近邻的思想实现向量转化和模型优化,从而得出网络节点表示向量,该过程模拟GCN的反向传播;最后,计算相似度矩阵,进行链路预测任务性能评测。在Citeseer数据集、DBLP数据集和Cora数据集上的实验结果表明:所提ALIP算法AUC值最高为98.01%,其性能优于其他23种链路预测算法,证明了该算法的有效性和可行性,同时也为无标签的复杂网络链路预测任务提供了一种新的解决方案。For unlabeled networks,the link prediction method based on graph neural networks had poor performance when using its efficient modeling mechanism for link prediction tasks.An unsupervised link prediction algorithm(ALIP)was proposed.It could approximate the graph neural network framework to simulate the efficient modeling mechanism and learning process of graph neural network algorithms,and to solve the problem of insufficient modeling caused by missing network node labels.Firstly,referring to the input layer of GCN,the structural information and node attributes of the network were fused.Secondly,matrix factorization is used to replace the hidden layer of GCN and simulate forward propagation.Then the ideas of identity mapping and high-order vector optimization to achieve vector transformation and model optimization to obtain the network node representation vector,which were used to simulate the back propagation of GCN.Finally,the similarity matrix for performance evaluation of link prediction tasks was calculated.On the Citeseer dataset,DBLP dataset and Cora dataset,the experimental results showed that ALIP algorithm had a maximum AUC value of 98.01%,and its performance was superior to the other 23 link prediction algorithms.The effectiveness and feasibility of the algorithm,in this study provide a new solution for complex unlabeled network link prediction tasks.

关 键 词：矩阵分解 向量优化 图卷积神经网络 相似度矩阵 链路预测 高阶近邻 

分 类 号：TP393.0[自动化与计算机技术—计算机应用技术]