融合社区连接信息的网络嵌入方法  

Network embedding method based on community connection information

作　　者：宋振寰 胡军[1,2] SONG Zhenhuan;HU Jun(Chongqing Key Laboratory of Computational Intelligence,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;School of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China)

机构地区：[1]重庆邮电大学计算智能重庆市重点实验室,重庆400065 [2]重庆邮电大学计算机科学与技术学院,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2023年第3期493-504,共12页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：国家自然科学基金(61772096,61876201,61876027);重庆市自然科学基金(cstc2019jcyj-cxttX0002,cstc2021ycjh-bgzxm0013);重庆市教委重点合作项目(HZ2021008)。

摘　　要：网络嵌入旨在学习节点的低维稠密向量,同时保留原始网络的结构和属性信息。现有的网络表示方法大多未考虑网络中的社区信息和社区间的信息,难以有效地学习网络的低维表示。为有效保留网络中的社区信息和社区间信息,提出了一种融合社区连接信息的网络嵌入方法(network embedding based on community connection information,ECCI)。该方法基于不同社区的亲密程度,捕捉网络中社区间的关系;采用自定义游走的方式得到融合局部结构、社区信息以及社区间信息的游走序列;通过Skip-Gram模型得到与之对应的网络嵌入结果。在3个公开数据集的实验结果表明,ECCI相比基准方法在链接预测上的AUC值和F1-Score都有一定程度的提升。Network embedding aims to learn the low-dimensional dense vector of nodes while preserving the structure and attribute information of an original network.Most of the existing network representation methods do not consider the community information and the connection information between communities,so it is difficult to effectively learn the low-dimensional representation of networks.In order to preserve community information and intercommunity information effectively,a network embedding based on community connection information(ECCI)is proposed to integrate community connection information.Firstly,the relationship between communities in a network is captured based on the intimacy degree of different communities.Then,the custom walk method is used to obtain the walking sequence of the local structure,community information and inter-community information of the network.Finally,the corresponding network embedding result is obtained by Skip-Gram model.Experimental results on three open data sets show that ECCI has a certain degree of improvement in AUC and F1-score of link prediction compared with the benchmark methods.

关 键 词：网络嵌入 Skip-Gram模型 社区发现 社区连接 链接预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]