基于高阶邻域信息交互的自监督异质图嵌入算法

打开文本图片集

关键词：异质图；自监督算法；节点嵌入；高阶邻域

中图分类号：TP391 文献标志码：A 文章编号：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

近年来，由于图嵌入[1，2]在分析图结构数据过程中十分重要，导致图嵌人技术发展十分迅速。（剩余18124字）