AGCFN:基于图神经网络多层网络社团检测模型  

AGCFN: multiplex network community detection modelbased on graph neural network

作　　者：陈龙 张振宇[2] 李晓明 白宏鹏 Chen Long;Zhang Zhenyu;Li Xiaoming;Bai Hongpeng(College of Software,Xinjiang University,rümqi 830000,China;College of Information Science&Engineering,Xinjiang University,rümqi 830000,China;College of International Business,Zhejiang Yuexiu University,Shaoxing Zhejiang 312000,China;College of Intelligence&Computing,Tianjin University,Tianjin 300000,China)

机构地区：[1]新疆大学软件学院,乌鲁木齐830000 [2]新疆大学信息科学与工程学院,乌鲁木齐830000 [3]浙江越秀外国语学院,浙江绍兴312000 [4]天津大学智能与计算学院,天津300000

出　　处：《计算机应用研究》2024年第10期2926-2931,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(62272311);国家重点研发计划资助项目(2018YFC0831005);中国天津经济技术开发区科技支撑计划资助项目(STCKJ2020-WRJ);中国新疆建设兵团第十二师财务科技项目(SR202103)。

摘　　要：基于图神经网络的多层网络社团检测方法面临以下两个挑战。一是如何有效利用多层网络的节点内容信息,二是如何有效利用多层网络的层间关系。因此,提出多层网络社团检测模型AGCFN(autoencoder-enhanced graph convolutional fusion network)。首先通过自编码器独立提取每个网络层的节点内容信息,通过传递算子将提取到的节点内容信息传递给图自编码器进行当前网络层节点内容信息与拓扑结构信息的融合,从而得到当前网络层每个节点的表示,这种方法充分利用了网络的节点内容信息与拓扑结构信息。对于得到的节点表示,通过模块度最大化模块和图解码器对其进行优化。其次,通过多层信息融合模块将每个网络层提取到的节点表示进行融合,得到每个节点的综合表示。最后,通过自训练机制训练模型并得到社团检测结果。与6个模型在三个数据集上进行对比,ACC与NMI评价指标有所提升,验证了AGCFN的有效性。Multiplex network community detection methods based on graph neural network face two main challenges.Firstly,how to effectively utilize the node content information of multiplex network;and secondly,how to effectively utilize the interlayer relationships in multiplex networks.Therefore,this paper proposed the multiplex network community detection model AGCFN.Firstly,the autoencoder independently extracted the node content information of each network layer and passed the extracted node content information to the graph autoencoder for fusing the node content information of the current network layer with the topology information through the transfer operator to obtain the representation of each node of the current network layer,which made full use of the node content information of the network and the topology information of the network.The modularity maximization module and graph decoder optimized the obtained node representation.Secondly,the multilayer information fusion module fused the node representations extracted from each network layer to obtain a comprehensive representation of each node.Finally,the model under went training,and it achieved community detection results through a self-training mechanism.Comparison with six models on three datasets demonstrate improvements in both ACC and NMI evaluation metrics,thereby va-lidating the effectiveness of AGCFN.

关 键 词：多层网络 社团检测 图神经网络 自编码器 自监督学习 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]