基于动态图表示学习的轻量化节点分类方法

打开文本图片集

引用格式：，等.基于动态图表示学习的轻量化节点分类方法[J].现代电子技术，2025,48（18)：1-8

关键词：动态图；节点分类；图表示学习；分组查询注意力机制；图神经网络；GAM模块中图分类号：TN911.72-34；TP391 文献标识码：A 文章编号：1004-373X（2025）18-0001-08

Method of lightweight node classification based on dynamic graph representation learning

YANQinyu,YANJinghua,BU Fanliang,WANG Yuzhe (CollegeofInformationandCyberSecurity,People'sPublicSecurityUniversityofChina,Beijing1Ooo38,China)

Abstract:Dynamic graph node classification isaclasicdownstream task in the fieldof graphrepresentation learning, aiming topredictthecategoriesofunlabelednodesbymeansofexisting informationindynamicgraphs.However,theexisting dynamicgraphnodeclasification methodsgenerallyhave theproblemofcomputational pressurecausedbythelargescaleand complex structureof themodel.Onthisbasis,amethodoflightweightnodeclasificationbasedondynamicgraphrepresentation learning(LNDG)isproposed.Inthis method,graphencoderisusedtoencodenode,linkandtimeinformationofdynamicgraph, andtheinovativeGAMmoduleisintroduced,whichutilizesthegrouped-queryatention(GQA)mechanismandtheMLP-Mixer method tofurtherlearnrepresentationsinboth temporalandspatialdimensions,achievingcompletecaptureof dynamicgraph features.TheexperimentalresultsonthreepublicclasicdatasetsshowthatLNDG hasanoverallparametersizeofonly0.70MB, which hasbeterAUCvaluesthan thebaseline model,andhas theadvantagesoflightweightand high eficiency.The proposed methodachievesagoodbalanceintermsofoverallsaleandfialefect,andhasgoodcomprehensiveperformanceinthetaskof dynamic graph node classification.

Keywords:dynamic graph;node clasification;graph representation learning;grouped-queryatention mechanism;graph neural network; GAM module

0 引言

节点分类是图表示学习的重要应用之一，受到研究人员的高度关注。（剩余18116字）