基于霍克斯过程的动态异质网络表征学习方法  

Dynamic heterogeneous network representation learning method based on Hawkes process

作　　者：陈蕾 邓琨[1,3] 刘星妍 CHEN Lei;DENG Kun;LIU Xingyan(College of Information Science and Engineering,Jiaxing University,Jiaxing 314001,China;School of Computer Science and Technology(School of Artificial Intelligence),Zhejiang Normal University,Jinhua 321004,China;Key Laboratory of Medical Electronics and Digital Health of Zhejiang Province,Jiaxing University,Jiaxing 314001,China)

机构地区：[1]嘉兴大学信息科学与工程学院,浙江嘉兴314001 [2]浙江师范大学计算机科学与技术学院(人工智能学院),浙江金华321004 [3]嘉兴大学浙江省医学电子与数字健康重点实验室,浙江嘉兴314001

出　　处：《电信科学》2024年第8期78-93,共16页Telecommunications Science

基　　金：教育部人文社会科学研究专项任务项目(No.22JDSZ3023);教育部产学合作协同育人项目(No.220603372015422,No.220604029012441)。

摘　　要：现有的异质网络表征学习方法主要关注静态网络,忽略了时间属性对节点表示的重要影响。然而,真实的异质信息网络极具动态性,节点和边的微小变化都可能影响整个结构和语义。鉴于此,提出了基于霍克斯过程的动态异质网络表征学习方法。首先,利用关系旋转编码方式和注意力机制,学习相邻节点的注意力系数,获得节点的向量表示。其次,学习不同元路径的最优加权组合以更好捕获网络的结构和语义信息。最后,基于时间衰减效应,通过邻域形成序列将时间特征引入节点表示中,得到节点的最终嵌入表示。在多种基准数据集上的实验结果表明,所提方法在性能上显著优于对比模型。在节点分类任务中,Macro-F1平均提高了0.15%~3.45%,在节点聚类任务中,归一化互信息(normalized mutual information,NMI)值提高了1.08%~3.57%。Existing methods for heterogeneous network representation learning mainly focus on static networks,overlooking the significant impact of temporal attributes on node representations.However,real heterogeneous informa‐tion networks are very dynamic,and even minor changes in nodes and edges can affect the entire structure and semantics.In this context,a dynamic heterogeneous network representation learning method based on Hawkes process was proposed.Firstly,the vector representation of nodes was obtained by utilizing the relational rotation encoding method and attention mechanism,where the attention coefficients of adjacent nodes were learned.Secondly,the optimal weighted combination of different meta-paths was learned to better captures the structural and semantic information of the network.Finally,leveraging the time decay effect,time features were introduced into node representations through the formation of neighborhood sequences,resulting in the ultimate embedding representation of nodes.Experimental results on various benchmark datasets indicate that the proposed method significantly outperforms baseline methods.In node classification tasks,Macro-F1 average is increased by 0.15%to 3.45%,and NMI value in node clustering tasks is improved by 1.08%to 3.57%.

关 键 词：网络表征学习 动态异质信息网络 注意力机制 元路径 霍克斯过程 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]