高级检索
    任嘉睿, 张海燕, 朱梦涵, 马波. 基于元图卷积的异质网络嵌入学习算法[J]. 计算机研究与发展, 2022, 59(8): 1683-1693. DOI: 10.7544/issn1000-1239.20220063
    引用本文: 任嘉睿, 张海燕, 朱梦涵, 马波. 基于元图卷积的异质网络嵌入学习算法[J]. 计算机研究与发展, 2022, 59(8): 1683-1693. DOI: 10.7544/issn1000-1239.20220063
    Ren Jiarui, Zhang Haiyan, Zhu Menghan, Ma Bo. Embedding Learning Algorithm for Heterogeneous Network Based on Meta-Graph Convolution[J]. Journal of Computer Research and Development, 2022, 59(8): 1683-1693. DOI: 10.7544/issn1000-1239.20220063
    Citation: Ren Jiarui, Zhang Haiyan, Zhu Menghan, Ma Bo. Embedding Learning Algorithm for Heterogeneous Network Based on Meta-Graph Convolution[J]. Journal of Computer Research and Development, 2022, 59(8): 1683-1693. DOI: 10.7544/issn1000-1239.20220063

    基于元图卷积的异质网络嵌入学习算法

    Embedding Learning Algorithm for Heterogeneous Network Based on Meta-Graph Convolution

    • 摘要: 异质网络嵌入是将异质网络中丰富的结构和语义信息嵌入到低维的节点表示中.图卷积网络是处理网络数据的一种有效方法,当前也被用于研究异质网络的多类型节点和多维关系的表示问题,现有的图卷积网络模型主要采用元路径来表示不同类型节点间的一种语义关系.然而,孤立的单条元路径无法准确地反映节点间的复杂语义,即不能充分利用节点间存在的多种高阶间接语义关系.针对上述问题,提出了一种基于元图卷积的异质网络嵌入学习算法MGCN(meta-graph convolutional network),包括基于元图的异构邻接矩阵计算以及学习节点的嵌入表示2个阶段,基于元图的异构邻接矩阵设计了融合多条元路径上的不同语义的计算方法,能够挖掘节点间的高阶间接关系,通过异构邻接矩阵的计算,能够聚合节点邻域特征为统一模式,此种卷积学习降低了图卷积方法的嵌入维数,从而减少了计算时间.在2个公开的异质网络数据集上进行社会计算基础研究任务的实验表明,MGCN在节点分类、聚类任务上比基线模型有更好的性能且需更少的训练时间.

       

      Abstract: Heterogeneous network embedding is to embed the rich structural and semantic information of heterogeneous networks into the low dimensional node representations. Graph convolutional networks are effective methods to process network data, and they are also used to research the representation of multi-type nodes and multi-dimensional relationships of heterogeneous networks. The existing graph convolutional network models mainly use meta-path to represent semantic relationship between nodes with different types. However, a single meta-path cannot accurately characterize the specific complex semantics between nodes, that is, it cannot make full use of high-order indirect semantic relationship between nodes. To address the above limitations, it is proposed that an embedding learning algorithm for heterogeneous network, named MGCN(meta-graph convolutional network). The algorithm includes two stages of heterogeneous adjacency matrices calculation based on meta-graph and learning node embedding. The heterogeneous adjacency matrix fuses different semantic information from multiple meta-paths and mines high-order indirect relationship between nodes. In addition, it can aggregate the neighborhood features of nodes into a unified pattern. This method reduces the embedding dimension, and then reduces the calculation time. Extensive experiments on two public heterogeneous network datasets show that the proposed MGCN can outperform baselines in basic research tasks of social computing like node classification and need less model training time.

       

    /

    返回文章
    返回