基于跨图特征融合和结构感知注意力的图相似度计算

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）08-010-2320-09

doi：10.19734/j.issn.1001-3695.2025.01.0020

Graph similarity computation based on cross-graph feature fusion and structure-aware attention

Pang Jun 1，2 ，Yan Bingxin’，LinXiaoli1，²，Wang Mengxiang3+ （1.ColegeofomputerScience&Tchnolog，WuhanUniersityofSiece&Techolog，Wuhan43O5，hina;2.HubeiProceKeyL boratoryfelos; jing 100088，China）

Abstract：GED isacommonlyused graph similarity metric function whose exact computation isanNP-hard problem.Therefore，recentlyresearchershave proposed numerousgraph neural network-based graph similaritycomputation methods.The existing methods ignrethecrossgraph interactioninformationbetweentwograph nodesduringfeature extractionandlack the learning of higher-orderrelationships betwee nodes inthe graph.Toaddresstheabove problems，this paperproposedamodel for graph similaritycomputationbasedoncross-graphfeature fusionand structure-aware atention.Firstly，themodel proposed across-graphnodefeature leaing method，which introducedacrossgraphattentiomechanismtoextractthecro-raphinteractioninformationof nodes，andeffectivelyfusedthelocal featuresofnodesandthecross-graphinteraction features.Secondly，he modelproposedastructure-awaremulti-atentionmechanism，whichcombinedthefeatureinformationofnodes withthegraph structural informationtoeficientlycapturethehigher-orderrelationshipsamong nodes.Experimentalesultson three public datasets show that the prediction accuracy of the CFSA model is improved by 4.8% ，5. 1% ，and 15.8% ， respectively，compared tothe existing models，and hasadvantages inalarge numberof performance metrics，which proves the effectiveness and efficiency of CFSA for the GED prediction task.

Key words： graph edit distance（GED）； graph similarity；graph neural network（GNN） ； graph embedding learning

0 引言

图数据因其独特的表达能力，能够有效地表示实体及实体之间的复杂关系，因此在推荐系统[1、社交网络分析和生物信息学[3]等领域得到广泛应用。（剩余19476字）