基于高階鄰域信息交互的自監(jiān)督異質(zhì)圖嵌入算法
打開文本圖片集
關(guān)鍵詞：異質(zhì)圖；自監(jiān)督算法；節(jié)點嵌入；高階鄰域
中圖分類號：TP391 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）07-012-2011-07
doi：10.19734/j. issn.1001-3695.2024.11.0493
Abstract：Toaddress theissue thatcurrentself-supervised neuralnetwork algorithms donotconsider theimpactofhighorder node information whenobtaining neighborhood node weights，thispaper proposedaself-supervised heterogeneous graph embeddingalgorithmbasedonhigh-orderneighborhood informationinteraction（SSHGEA-HNI）.Itenhancedlocaloptimization capabilitiesandmodelperformancebyaddingafeedforwardfullconnectedlayerintheattentionmechanism tocapturehighorderneighborhoodnodefeatures.Thealgorithmconsistedofalabel generationmoduleandanembeddinglearning module.The label generation module produced pseudo-labels for nodesthrough label propagation，which servedassupervisorysignals to guidetheembedding generationmodule to produceembeddings.Theembedding learning module generatedembeddings andattentioncoeffcientsthroughtheatentionmechanismbasedonhighorderneighborhoodinformationinteraction，withtheaentioncoeffcientsusedtoguidethelabelgeneration module toproduce pseudo-labels.Ineach iteration，thetwo modules shared node atentioncoeficients，promoting mutual utilizationandenhancementbetweenthetwo modules.Experimentswereconductedonfourreal heterogeneous graphdatasets，withimprovementsobserved intheclusteringand clasification tasksof most datasets.Theexperimental results demonstrate thatthe modelcan efectivelyutilize high-order node information.
Key words：heterogeneous graph；self-supervised algorithm；node embedding；high-order neighborhood
0 引言
近年來，由于圖嵌入[1，2]在分析圖結(jié)構(gòu)數(shù)據(jù)過程中十分重要，導(dǎo)致圖嵌人技術(shù)發(fā)展十分迅速。（剩余18124字）