一种基于相似性注意力的社区发现模型  

A Community Detection Model Based on Similarity Attention

作　　者：杜航原[1] 姚倩 冯劲松 王文剑[1] DU Hangyuan;YAO Qian;FENG Jinsong;WANGWenjian(College of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Taiyuan Sanitation Industry Management Center,Taiyuan 030002,China)

机构地区：[1]山西大学计算机与信息技术学院,山西太原030006 [2]太原市环卫产业管理中心,山西太原030002

出　　处：《山西大学学报（自然科学版）》2022年第6期1513-1522,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61902227,62076154,U1805263,61773247);山西省自然科学基金(201901D211192);山西省高等学校科技创新项目(2019L0039)。

摘　　要：图卷积网络可以有效融合属性网络的拓扑结构和节点属性,已被应用于社区发现领域中。然而,现实中的复杂网络规模庞大,导致图卷积网络无法识别出对社区结构的形成具有较大影响力的节点,从而带来严重的噪声。针对这一问题,提出了一种基于相似性注意力的社区发现模型。该模型以编码-解码方式构建信息传播框架,将图注意力网络作为编码器对网络结构和节点属性进行深度融合,使用相似性注意力为影响力更大的节点赋予更高的权重,并设计了结构解码器和属性解码器分别用于重构网络的拓扑信息和属性信息;构造模块度优化器实现社区结构的划分与优化。在真实网络数据集上对本文提出模型的有效性进行了验证,结果表明该模型相比Infomap、LPA、GAE和JGE-CD等方法具有更好的社区发现能力。Graph convolutional network can effectively fuse topology structure and attribute information of nodes,and has been applied in the field of community detection.However,due to the large scale of complex networks in reality,the graph convolution network cannot identify the nodes that have a great influence on the formation of community structure,resulting in serious noise.Aiming at the problem,a community discovery model based on similarity attention was proposed.The model was constructed based on an information dissemination frame in an encoding-decoding manner,and it fullfils deep integration of the network structure and node attibutes by using the graph attention network as the encoder,in which the high weights of more influential nodes are assigned using similartity attention.The effectiveness of the proposed model is verified on real network datasets,and the results show that the model has better performance in community detection than Infomap,LPA,GAE and JGE-CD methods.

关 键 词：社区发现 属性网络 图注意力网络 相似性注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]