基于多粒度的异质图表示  

Heterogeneous Graph Representations Based on Multi-granularity

作　　者：芮品德 赵桓幜 赵姝[1,2,3] 张燕平 RUI Pinde;ZHAO Huanjing;ZHAO Shu;ZHANG Yanping(Key Laboratory of Intelligent Computing and Signal Processing,Ministry of Education,Hefei 230601,China;School of Computer Science and Technology,Anhui University,Hefei 230601,China;Information Materials and Intelligent Sensing Laboratory of Anhui Province,Hefei 230601,China)

机构地区：[1]计算与信号处理教育部重点实验室,安徽合肥230601 [2]安徽大学计算机科学与技术学院,安徽合肥230601 [3]安徽省信息材料与智能传感重点实验室,安徽合肥230601

出　　处：《山西大学学报（自然科学版）》2023年第1期20-30,共11页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61876001);国家重点研发计划(2017YFB1401903);安徽省自然科学基金(1708085QF156)。

摘　　要：异质图表示学习旨在将图中的语义信息和异质的结构信息嵌入到低维向量空间中。目前大多数的异质图表示学习方法主要通过基于元路径、元图和网络模式的采样以保留图中同类型节点间的单粒度局部结构,忽略了现实世界中复杂异质图具有的丰富的层次结构。商空间理论中的多粒度思想可以在不同粒度内捕获节点间的潜在联系。因此,为在异质图表示中有效地保留层次结构的信息,文章提出一个基于多粒度的异质图表示方法(Heterogeneous Graph Representations Based on Multi-granularity,HeMug)。该方法首先基于不同元路径构建多个同质子图,并利用多粒度的粗化思想,将每个同质子图分别粗化形成多个多粒度子网络,以保留异质图中同类型节点在给定元路径下的层次结构。其次,利用多粒度的细化思想,将每个多粒度子网络最粗层通过现有表示学习方法获得的节点表示逐层细化,以得到节点在每个多粒度子网络下的表示。最后,设计注意力机制以融合节点在不同元路径对应的多粒度子网络下的表示。在四个真实数据集上的实验结果表明,与对比算法相比,提出的HeMug获得了更有效的节点表示。Heterogeneous graph representation learning aims to embed semantic and structural information from heterogeneous graphs into a low-dimensional vector space. Most current heterogeneous graph representation learning methods are mainly based on sampling of meta-path, meta-graph and network schema to preserve the single-grained local structure among nodes of the same type in the graph, however, ignore the rich hierarchical structure of complex heterogeneous graphs in the real world. The idea of multigranularity coarsening and refinement can capture the potential relationships between nodes at different granularities. Therefore, in order to effectively retain the information of the hierarchical structure in the heterogeneous graph representation, we have proposed a multi-granularity based learning method for heterogeneous graph representation, called HeMug. HeMug first constructs multiple homogeneous subgraphs based on different meta-paths, and utilizes the coarsening idea of multi-granularity to coarsen each homogeneous subgraph separately to form multiple multi-granularity subnetworks, thus preserves the hierarchical structure of homogeneous nodes in heterogeneous graphs under a given meta-path. Second, the node representation of the coarsest layer of each multigranularity subnetwork is obtained by existing representation learning methods, and utilizing the refinement idea of multi-granularity, the node representation of the coarsest layer is refined layer by layer to obtain the representation of the node under each multi-granularity sub-network. Finally, the attention mechanism is designed to fuse the representation of nodes under a multi-granularity network corresponding to different meta-paths. Experimental results on four real datasets show that the HeMug obtains a more efficient node representation compared to the comparison algorithm.

关 键 词：异质图 图表示 多粒度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]