【作者】 陈伟；

【导师】 周丽华；

【作者基本信息】 云南大学 ， 计算机应用技术， 2023， 硕士

【摘要】 异质信息网络(Heterogeneous Information Network,HINs)中的社区搜索旨在寻找一组包含查询节点且密切相关的同类型节点集合,为用户提供了一种查询信息的方式,有助于好友推荐、疫情监控以及蛋白质功能预测等应用,近年来受到了工业界和学术界的广泛关注。然而现有HINs社区搜索方法大多基于预定义的子图模式对社区的拓扑结构施加一个严格的要求,忽略了节点间的属性相似性,导致结构关系弱而属性相似性高的社区难以定位,并且采用的全局搜索模式,难以有效处理大规模的网络数据。此外,基于图神经网络模型的社区搜索方法虽然实现了具有规模约束的结构内聚且属性相似的社区搜索,但现有方法都是针对单一类型的同质网络而设计的,不适用于包含多种类型节点和连边的异质信息网络;并且现有的图神经网络模型都不能在区分邻域形成的异质性因素的同时,充分抽取和融合HINs的语义信息,进而不能很好地用于处理HINs中的社区搜索问题。针对上述问题,本文进行了异质信息网络中基于解耦图神经网络的社区搜索研究,本文的主要工作如下:(1)提出了具有社区规模约束的属性异质信息网络社区搜索问题,旨在异质信息网络中搜索结构凝聚、属性相似、节点数目不超过c的同类型节点集合。首先设计基于元路径的局部模块度用于度量节点间的结构内聚性,搭建解耦图神经网络(Disentangled graph neural network,DGNN),通过学习节点表征推算节点间的属性相似性,然后利用0/1背包问题优化属性和结构两种凝聚性度量指标,定义了最有价值的c大小社区搜索问题,进而提出了一种基于解耦图神经网络的价值最大化社区搜索模型(Value Maximization Community Search Based on Disentangled Graph Neural Network,VMCS-DGNN),VMCS-DGNN在通过构造候选子图保证搜索效率的基础上,实现了 HINs中属性相似且结构内聚的社区搜索。(2)设计了异质解耦图神经网络模型(Heterogeneous Disentangled Graph Neural Network,HDGNN),进而提出了基于HDGNN的价值最大化社区搜索模型(Value Maximization Community Search Based on Heterogeneous Disentangled Graph Neural Network,VMCS-HDGNN)。HDGNN将单通道的解耦图卷积网络DGNN扩展为多通道的解耦图卷积网络,通过使用两级解耦图卷积网络,在区分邻域形成的异质性因素的同时,充分抽取与融合了 HINs丰富的语义信息,进而提高节点间属性相似度计算的准确性。(3)在六个(三个同质和三个异质)真实的网络数据集上,通过不同搜索方法搜索结果的对比、消融和参数分析等实验,验证了 VMCS-DGNN模型的有效性和高效性;在三个真实的异质信息网络数据集上,通过社区搜索性能评估、消融实验和参数敏感性分析,对VMCS-HDGNN的有效性和稳定性进行评估。实验结果验证了所提模型均优于基线模型,其中VMCS-HDGNN在各个数据集上的社区搜索准确率比VMCS-DGNN均提升了 6%以上,证明了 HDGNN能够充分融合HINs丰富的语义信息。更多还原

【Abstract】 The purpose of community search over heterogeneous information networks is to find a group of closely related nodes of the same type that contain query nodes.It providing users with a way to query information,which is helpful for friend recommendation,epidemic monitoring,protein function prediction and other applications.In recent years,it has received widespread attention from academia and industrial.However,most of the existing community search methods over HINs impose strict requirements on the topology of the community based on the predefined subgraph pattern,ignoring the attribute similarity between nodes,which will be difficult to locate the community with weak structural relationship and high attribute similarity.And the global search mode is difficult to effectively deal with large-scale network data.In addition,community search methods that use graph neural network model have achieved community search with structure cohesion and similar attributes,but existing methods are designed for homogeneous networks that contain only one node type and one connection type.They are not suitable for heterogeneous information networks that contain multiple types of nodes and connections.Moreover,the existing graph neural network models can’t fully extract and fuse the semantic information of HINs when distinguishing the heterogeneous factors formed by the neighborhood,and can’t be well applied to deal with the community search problem over heterogeneous information networks.In view of the above problems,this thesis conducted research on community search over heterogeneous information networks based on decoupling graph neural networks.The main work of this thesis is as follows:(1)This thesis propose a community search problem over attribute heterogeneous information networks with community size constraints.It aiming at searching the same type of node set with condensed structure,and similar attributes and no more than c nodes in heterogeneous information networks.First,this thesis design a local modular degree based on meta-path to measure the structural cohesion between nodes,build a disentangled graph neural network DGNN,and calculate the attribute similarity between nodes through learning node representation.Then,this thesis define the most valuable csize community search problem by using the 0/1 knapsack problem to optimize the attribute and structure two cohesion metrics.Furthermore,this thesis propose a value maximization community search model VMCS-DGNN based on disentangled graph neural network.The method VMCS-DGNN achieves community search with similar attributes and cohesive structure over HINs while ensuring search efficiency by constructing candidate subgraphs.(2)This thesis designed the heterogeneous disentangled graph neural network model HDGNN,and then this thesis proposed the value maximization community search model VMCS-HDGNN.HDGNN extends the single-channel disentangled graph convolution network DGNN to the multi-channel disentangled graph convolution network.The model HDGNN uses a two-level disentangled graph convolutional network to fully extract and fuse the rich semantic information over HINs while distinguishing the heterogeneity factors formed by the neighborhood,thereby improving the accuracy of attribute similarity calculation between nodes.(3)This thesis have verified the validity and efficiency of VMCS-DGNN model through comparison of different search methods,ablation experiments and parameter analysis on six(three homogeneous and three heterogeneous)real network data sets.This thesis evaluated the effectiveness and stability of VMCS-HDGNN on three real heterogeneous information network data sets through community search performance evaluation,ablation experiment and parameter sensitivity analysis.The experimental results verified that the proposed model is superior to the baseline model.The community search accuracy of VMCS-HDGNN in all data sets has improved by more than 6%compared with VMCS-DGNN,which proves that HDGNN can fully integrate the rich semantic information of HINs.更多还原