基于相似网络和联合注意力的图嵌入模型
摘要:
图注意力网络(graph attention network,GAT)将注意力机制与图神经网络融合,但模型只关注节点的一阶邻域节点,缺乏对高阶相似节点的考虑,同时在计算注意力分数时缺乏对节点结构特征的关注。为此提出一种基于相似网络和联合注意力的图嵌入模型。首先计算网络中的节点相似性,并将高相似度且未连接的节点对构建新边以形成相似网络。其次,引入结构相关性和内容相关性的概念,分别用 于表征节点之间的结构关系和内容特征。通过融合两种相关性得分计算得到联合注意力分数。最后使用联合注意力分数对节点特征加权聚合,得到最终的节点嵌入表示。将本文所提算法在Cora、Citeseer和Pubmed3个数据集上进行节点分类任务,准确率分别达到85.70%、74.30% 、84.10% ,与原始图注意力网络模型相比分别提高了2.70% 、3.94%和2.60%。可见 ,所提出的算法可以得到更好的节点嵌入表示 。
The Graph Attention Network(GAT)incorporates the attention mechanism into graph neural networks. However,the model only considers the first-order neighborhood nodes of nodes,neglecting the consideration of higher-order similar nodes,and fails to account for the structural features of nodes when calculating the attention score.To address this is-sue,this paper propose a graph embedding model based on higher-order similar nodes and joint at tention.Specifically,our ap-proach first computes node similarities in the network and subsequently constructs new edges between pairs of nodes that are highly similar but not directly connected,thus forming a similar network.Secondly,we introduce the notions of structural rele-vance and content relevance to respectively characterize structural relationships and content features among nodes.Finally,we perform weighted aggregation of the node features using joint attention scores to obtain the final node embedding representations.The accuracy improvement over traditional models was found to be 2.70%,3.94%,and2.60%,respectively.These re-sults demonstrate that the proposed method yields a better representation of node embedding.
作者:
王静红,李昌鑫,杨家腾,于富强
Wang Jinghong,Li Changxin Yang Jiateng,Yu Fuqiang
机构地区:
河北师范大学计算机与网络空间安全学院;河北省网络与信息安全重点实验室;供应链大数据分析与数据安全河北省工程研究中心;河北工程技术学院人工智能与大数据学院
引用本文:
王静红,李昌鑫,杨家腾等。基于相似网络和联合注意力的图嵌入模型[J].betway官方app 学报(自然科学版),2024,52(6) :36-44.
(WangJinghong,LiChangxin,YangJiateng,etal.A graph embedding modelbased on simi- larnetworks andjointattention[J] .JournalofHenan NormalUniversity(NaturalScienceEdition) ,2024,52(6):36-44. DOI:10. 16366/j. cnki.1000-2367. 2023. 06. 16. 0001. )
基金:
河北省自然科学基金;河北省高等学校科学技术研究项目;中央引导地方科技发展资金项目;河北省归国人才资助项目
关键词:
图嵌入;图注意力网络;节点相似性;相似网络;节点分类
graph embedding;graph attention network;node similarity;similar network;nodec lassification
分类号:
TP181