基于门控循环单元的移动社会网络链路预测方法刘林峰于子兴祝贺（南京邮电大学计算机学院南京210023）（956756009@qq.com）ALinkPredictionMethodBasedonGatedRecurrentUnitsforMobileSocialNetworkLiuLinfeng,YuZixing,andZhuHe（CollegeofComputerScience,NanjingUniversityofPostsandTelecommunications,Nanjing210023）AbstractLinkpredictionisdefinedasthepredictionofpotentialrelationshipsbetweennodesinthefuturebasedontheknownnetworktopologyandthenodeinformation.Linkpredictioncanhelpreduceresourceexpenditureandallocateresourcesmorereasonablyinvariousapplicationsincludinglinks.Mobilesocialnetworkisakindofdynamicnetwork,anditsstructureisalwaysevolvingwiththeappearanceanddisappearanceofnodesandlinksovertime.Accordingtothecharacteristicsofthemobilesocialnetwork,thecurrentexistingresearchesusemoresophisticatedmodeltoanalyzetherelationshipbetweenthelinks,howevercomplexmodelsnotonlyhavelargespacecomplexitybutalsoareeasytooverfittingproblem.Inordertosolvetheaboveproblems,agatingcycleunitbasedonthepredictionmethodofmobilesocialnetworklinkisputforward.Firstly,theinputdatasetissortedandfiltered,andthetargetnetworkisdividedintosnapshotgraphsandtransformedintoadjacencymatricestoformasampleset.Then,thepredictionmodelisconstructedbasedontheautoencoderandthegatedrecurrentunitstoextractthetemporalcharacteristicsofmobilesocialnetwork.BasedonKONECTdataset,theexperimentalresultscomparedwithothermodelsshowthattheproposedmethodcanimprovethetrainingefficiencyby49.81%,whilethepredictionperformancecanbemaintained.Keywordsdeeplearning；linkprediction；mobilesocialnetwork；gatedrecurrentunits；autoencoder摘要链路预测是指通过已知的网络拓扑和节点信息来预测未来时刻节点之间的潜在关系，链路预测能够帮助在各种存在链路的应用领域更加合理地分配资源、降低资源开销.移动社会网络属于动态网络的一种，其网络结构总是随着节点和链路的出现、消失以及时间推移而不断演变.针对移动社会网络的特点，当前已有的研究使用愈加复杂的模型来分析链路之间的联系，然而复杂的模型不但空间复杂度大而且容易造成过拟合问题.为了解决以上问题，提出一种基于门控循环...
