融合路径优化的异构图神经网络算法  

Path Optimization Based Heterogeneous Graph Neural Network Algorithm

作　　者：秦志龙 朱一峰 邓琨 雍剑书 QIN Zhilong;ZHU Yifeng;DENG Kun;YONG Jianshu(School of Computer Science and Technology(School of Artificial Intelligence),Zhejiang Sci-Tech University,Hangzhou 310018,China;College of Information Science and Engineering,Jiaxing University,Jiaxing 314001,China;Jiaxing Hengchuang Power Group Co.,Ltd.Jiachuang Comprehensive Service Branch,Jiaxing 314000,China)

机构地区：[1]浙江理工大学计算机科学与技术学院(人工智能学院),杭州310018 [2]嘉兴大学信息科学与工程学院,浙江嘉兴314001 [3]嘉兴恒创电力集团有限公司佳创综合服务分公司,浙江嘉兴314000

出　　处：《小型微型计算机系统》2025年第3期627-635,共9页Journal of Chinese Computer Systems

基　　金：教育部人文社会科学研究专项任务项目(22JDSZ3023)资助;教育部产学合作协同育人项目(220603372015422,220604029012441)资助.

摘　　要：图神经网络作为处理图结构数据的一种有效方法,可以有效抽取异构图中的复杂结构与语义信息,在节点分类和连接预测任务上取得了优异表现.然而现有异构图神经网络算法忽略元路径下各个节点类型之间的相关性,导致在语义融合、更新时丢失邻域结构特征信息,从而影响模型整体性能.为解决该问题,提出融合路径优化的异构图神经网络算法.首先用特征传播使所有类型节点获得属性特征;其次通过元路径实例得到节点中心性信息;随后采用优化相似度计算不同类型节点贡献程度,学习异构图语义信息;最后提出路径优化策略进行多层训练,捕获节点之间潜在关联,获得节点嵌入表示.在ACM、IMDB和DBLP数据集上进行广泛实验,并与当前主流算法进行对比分析,实验结果证明了该方法的有效性.As an effective method for processing graph structure data,graph neural networks can effectively extract complex structural and semantic information from heterogeneous graphs,and have achieved excellent performance in node classification and connection prediction tasks.However,the existing heterogeneous graph neural network algorithms ignore the correlation between each node type under the meta-path,resulting in the loss of neighborhood structure feature information during semantic fusion and updating,thus affecting the overall performance of the model.In order to solve the above problems,a heterogeneous graph neural network algorithm based on path optimization was proposed.We first use feature propagation to make all types of nodes obtain attribute features.Secondly,the node centrality information is obtained through the meta-path instances.Then,the optimization similarity is used to calculate the contribution degree of different types of nodes,and the heterogeneous graph semantic information is learned.Finally,the path optimization strategy which captures the potential association between nodes is proposed for multi-layer training and node embedding representation.Extensive experiments on ACM,IMDB and DBLP datasets compared with current mainstream algorithms,and the experimental results show that the proposed method is practical.

关 键 词：异构图 图神经网络 元路径 路径优化 图嵌入 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]