摘要
为解决异质网络重叠社区发现问题,提出一种基于异质图注意力网络的重叠社区发现模型。通过异质图注意力网络的双层注意力机制,从节点级与语义级充分挖掘节点、元路径在信息表示中的重要程度,并进行分层聚合获得节点特征向量,将伯努利-泊松模型与图卷积网络整合,在生成社区隶属关系矩阵的基础上优化社区重叠度进行重叠社区划分。模型改变异质图注意力网络的激活函数,改善梯度消失问题,将异质图注意力网络与伯努利-泊松模型结合实现异质网络的重叠社区发现。使用真实数据集进行实验,实验结果表明,模型可以利用异质网络节点信息多样性进行重叠社区发现,相对传统社区发现方法具备较好的稳定性和准确性。
To solve the problem of overlapping community discovery in heterogeneous networks,a model of overlapping community discovery was proposed based on heterogeneous graph attention network.A two-layer attention mechanism including the node level and the semantic level was built,the importance of nodes and meta-paths in information representation was learnt.Bernoulli-Poisson model was integrated with the graph convolutional neural network.Based on the generated community affiliation matrix,the community overlap degree of the network was optimized for the discovery of overlapping communities.It changed the activation function of the heterogeneous graph attention network to improve the gradient disappearance problem,and the heterogeneous graph attention network was combined with the Bernoulli-Poisson model to achieve overlapping community discovery in heterogeneous networks.Experiments were conducted using real data sets.Experimental results show that the model can fully consider the diversity of heterogeneous network node information for overlapping community discovery,and it is compared with traditional methods,the model has better stability and accuracy.
作者
孙悦
赵宇红
薛婷
SUN Yue;ZHAO Yu-hong;XUE Ting(School of Information Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,China)
出处
《计算机工程与设计》
北大核心
2023年第12期3649-3655,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(81571753)
内蒙古自然科学基金项目(2022MS06006)。
关键词
重叠社区发现
异质网络
异质图注意网络
元路径
图卷积神经网络
伯努利-泊松模型
社区隶属矩阵
overlapping community discovery
heterogeneous networks
heterogeneous graph attention network
meta-path
graph convolutional neural network
Bernoulli-Poisson model
community affiliation matrix
