Article(id=1329457284658344645, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1329457157424132411, articleNumber=null, orderNo=null, doi=10.7544/issn1000-1239.202440192, pmid=null, cstr=null, oa=null, hot=null, price=null, onlineType=0, articleFormat=0, articleType=null, articleTypeStr=null, receivedDate=1710518400000, receivedDateStr=2024-03-16, revisedDate=1715702400000, revisedDateStr=2024-05-15, acceptedDate=null, acceptedDateStr=null, onlineDate=1794896916784, onlineDateStr=2026-11-17, pubDate=1722441600000, pubDateStr=2024-08-01, doiRegisterDate=null, doiRegisterDateStr=null, onlineIssueDate=1794896916784, onlineIssueDateStr=2026-11-17, onlineJustAcceptDate=null, onlineJustAcceptDateStr=null, onlineFirstDate=null, onlineFirstDateStr=null, sourceXml=null, magXml=null, createTime=1794896916784, creator=13701087609, updateTime=1794896916784, updator=13701087609, issue=Issue{id=1329457157424132411, tenantId=1146029695717560320, journalId=1324759914146529478, year='2024', volume='61', issue='8', pageStart='1863', pageEnd='2126', issueExtLink='null', onlineDate='null', pubDate='1722441600000', pubDateStr='2024-08-01', beforeIssueId=null, nextIssueId=null, price=null, status=1, issueComplete=1, articleOrder=1, issueType=-1, specialIssue=null, createTime=1794896886449, creator='13701087609', updateTime=1794896886449, updator='13701087609', preIssue=null, nextIssue=null, articleTotal=null, ext=null, issueFiles=null, downloadFileDto=null}, startPage=2045, endPage=2066, ext={EN=ArticleExt(id=1329457286264763079, articleId=1329457284658344645, tenantId=1146029695717560320, journalId=1324759914146529478, language=EN, title=Survey on Graph Neural Ordinary Differential Equations, columnId=1329457158304936255, journalTitle=Journal of Computer Research and Development, columnName=Special Issue on Graph Machine Learning, runingTitle=null, highlight=null, articleAbstract=

Graph neural networks (GNNs) are powerful tools for handling graph-structured data, capable of capturing complex relationships and features among nodes. However, the discrete architecture of GNNs leads to numerous challenges in representing graph structures, modeling graph evolution, adapting to irregular data, and managing computational costs. In response to these challenges, neural ordinary differential equations (ODEs) have been introduced as a novel method to address the challenges faced by GNNs, as they can simulate the continuous evolution of system states, providing continuous deep encoding and inference capabilities. However, neural ODEs are designed for Euclidean structured data and cannot directly capture the characteristics of graphs. Therefore, researchers have proposed graph neural ODEs, a new type of architectures that combines neural ODEs with GNNs, which can better adapt to graph-structured data and fully utilize its characteristics. In recent years, research related to graph neural ODEs has delved into various directions of graph machine learning, sparking a new research trend. In this context, we systematically review the relevant research of graph neural ODEs in a timely manner. Firstly, we review the key advantages of GNNs and the challenges they face, and elucidate the theoretical basis and practical significance of introducing neural ODEs and combining them with GNNs. Subsequently, we elaborate on the background and basic concepts of graph neural ODEs, proposing a novel taxonomy, and comprehensively describe some important current methods on the taxonomy. Then, we introduce commonly used verification methods in related research, including downstream tasks and datasets. Furthermore, we delve into the applications of graph neural ODEs in multiple practical fields. Finally, we summarize and prospect the challenges and future development trends of graph neural ODEs.

, authors=Pengfei Jiao1, Shuxin Chen2, Xuan Guo3, Dongxiao He3, *, Dong Liu4, authorsList=Pengfei Jiao, Shuxin Chen, Xuan Guo, Dongxiao He, Dong Liu, authorCompany=null, correspAuthors=Dongxiao He, authorNote=

born in 1990. PhD, professor. His main research interest includes complex network analysis and its applications

born in 2004. Undergraduate. Her main research interest includes graph neural network

born in 1996. PhD candidate. His main research interest includes graph machine learning

born in 1984. PhD, professor. Her main research interests include analysis of complex networks, graph data mining, and graph neural network

born in 1976. PhD, professor. His main research interests include educational data mining and complex network analysis

, correspAuthorsNote=null, copyrightStatement=Copyright © 2024 Journal of Computer Research and Development. All rights reserved., copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=null, magXml=null, pdfUrl=null, pdf=null, pdfFileSize=null, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=null, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=null, mapNumber=null, fund=null), CN=ArticleExt(id=1329457288131228386, articleId=1329457284658344645, tenantId=1146029695717560320, journalId=1324759914146529478, language=CN, title=图神经常微分方程综述, columnId=1329457158523040065, journalTitle=计算机研究与发展, columnName=图机器学习专题, runingTitle=null, highlight=null, articleAbstract=

图神经网络(graph neural network,GNN)是处理图结构数据的强大工具,能够捕捉节点间的复杂关系和特征. 但GNN的离散架构导致其在表示图结构、建模图演化、适应不规则数据和计算开销等方面面临诸多挑战. 面对这些挑战,神经常微分方程(ordinary differential equation,ODE)由于能够模拟系统状态的连续变化,具备“连续深度”的编码和推断能力,被作为解决GNN面临挑战的全新方法而引入. 然而,神经ODE是为欧式结构数据设计的,无法直接捕捉图结构特性. 因此,提出了图神经ODE,一种将神经ODE与GNN结合的新架构,可以更好地适应图结构数据并充分利用其特性. 近年来,图神经ODE相关研究已经深入到图机器学习的各个方向中,引发了新的研究热潮. 在此背景下,适时地对图神经ODE研究前沿进行了系统性综述. 首先,回顾了GNN的关键优势和面临的诸多挑战,阐述了引入神经ODE并与GNN结合的理论基础和实践意义. 随后,详细介绍了图神经ODE的背景和基本概念,并提出了一种新颖的分类方法,在此基础上对当前的相关方法进行了全面描述. 然后,介绍了相关研究常用的验证方法,包括下游任务及数据集. 进一步,深入探讨了图神经ODE在多个实用领域上的应用. 最后,对图神经ODE面临的挑战和未来发展趋势进行了总结和展望.

, authors=焦鹏飞1, 陈舒欣2, 郭翾3, 何东晓3, *, 刘栋4, authorsList=焦鹏飞, 陈舒欣, 郭翾, 何东晓, 刘栋, authorCompany=null, correspAuthors=何东晓, authorNote=

1990年生. 博士,教授. 主要研究方向为复杂网络分析及其应用

2004年生. 本科生. 主要研究方向为图神经网络

1996年生. 博士研究生. 主要研究方向为图机器学习

1984年生. 博士,教授. 主要研究方向为复杂网络分析、图数据挖掘、图神经网络

1976年生. 博士,教授. 主要研究方向为教育数据挖掘、复杂网络分析

, correspAuthorsNote=
何东晓()
, copyrightStatement=版权所有 © 2024 计算机研究与发展编辑部, copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=+iikfS9v88hlBm5bU+iC2A==, magXml=GZsF1ZIsperYROt/n2bBOw==, pdfUrl=null, pdf=FUuy4UefemWtLCqKxtXP5A==, pdfFileSize=2142295, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=x9Yp13Vs2ygQxslfTGMNnA==, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=VZXkPUL1EGxUgnObYPtgHA==, mapNumber=null, fund=null)}, authors=[Author(id=1329457288680682226, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=pjiao@hdu.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457288760374004, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288680682226, language=EN, stringName=Pengfei Jiao, firstName=Pengfei, middleName=null, lastName=Jiao, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Cyberspace, Hangzhou Dianzi University, Hangzhou 310018, bio={"content":"

born in 1990. PhD, professor. His main research interest includes complex network analysis and its applications

"}, bioImg=null, bioContent=

born in 1990. PhD, professor. His main research interest includes complex network analysis and its applications

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457288840065781, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288680682226, language=CN, stringName=焦鹏飞, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1杭州电子科技大学网络空间安全学院 杭州 310018, bio={"img":"xkFJCUTEussLht1CvuM/7w==","content":"

1990年生. 博士,教授. 主要研究方向为复杂网络分析及其应用

"}, bioImg=xkFJCUTEussLht1CvuM/7w==, bioContent=

1990年生. 博士,教授. 主要研究方向为复杂网络分析及其应用

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288340943588, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=1, ext=[AuthorCompanyExt(id=1329457288349332197, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288340943588, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Cyberspace, Hangzhou Dianzi University, Hangzhou 310018), AuthorCompanyExt(id=1329457288357720806, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288340943588, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1杭州电子科技大学网络空间安全学院 杭州 310018)])]), Author(id=1329457288919757559, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457292392641274, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288919757559, language=EN, stringName=Shuxin Chen, firstName=Shuxin, middleName=null, lastName=Chen, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2ZhuoYue Honors College, Hangzhou Dianzi University, Hangzhou 310018, bio={"content":"

born in 2004. Undergraduate. Her main research interest includes graph neural network

"}, bioImg=null, bioContent=

born in 2004. Undergraduate. Her main research interest includes graph neural network

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457292489110268, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288919757559, language=CN, stringName=陈舒欣, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2杭州电子科技大学卓越学院 杭州 310018, bio={"img":"vecXV3jDToAN4du0WSeQOQ==","content":"

2004年生. 本科生. 主要研究方向为图神经网络

"}, bioImg=vecXV3jDToAN4du0WSeQOQ==, bioContent=

2004年生. 本科生. 主要研究方向为图神经网络

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288408052455, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=2, ext=[AuthorCompanyExt(id=1329457288412246760, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288408052455, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2ZhuoYue Honors College, Hangzhou Dianzi University, Hangzhou 310018), AuthorCompanyExt(id=1329457288420635369, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288408052455, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2杭州电子科技大学卓越学院 杭州 310018)])]), Author(id=1329457292560413438, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457292640105216, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292560413438, language=EN, stringName=Xuan Guo, firstName=Xuan, middleName=null, lastName=Guo, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3College of Intelligence and Computing, Tianjin University, Tianjin 300350, bio={"content":"

born in 1996. PhD candidate. His main research interest includes graph machine learning

"}, bioImg=null, bioContent=

born in 1996. PhD candidate. His main research interest includes graph machine learning

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457292723991297, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292560413438, language=CN, stringName=郭翾, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3天津大学智能与计算学部 天津 300350, bio={"img":"reHGELmAdLzAZJBAkdfn+Q==","content":"

1996年生. 博士研究生. 主要研究方向为图机器学习

"}, bioImg=reHGELmAdLzAZJBAkdfn+Q==, bioContent=

1996年生. 博士研究生. 主要研究方向为图机器学习

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288491938538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=3, ext=[AuthorCompanyExt(id=1329457288500327147, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Intelligence and Computing, Tianjin University, Tianjin 300350), AuthorCompanyExt(id=1329457288512910060, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3天津大学智能与计算学部 天津 300350)])]), Author(id=1329457292786905860, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=hedongxiao@tju.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1329457292858209030, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292786905860, language=EN, stringName=Dongxiao He, firstName=Dongxiao, middleName=null, lastName=He, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, *, address=3College of Intelligence and Computing, Tianjin University, Tianjin 300350, bio={"content":"

born in 1984. PhD, professor. Her main research interests include analysis of complex networks, graph data mining, and graph neural network

"}, bioImg=null, bioContent=

born in 1984. PhD, professor. Her main research interests include analysis of complex networks, graph data mining, and graph neural network

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457292954678023, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292786905860, language=CN, stringName=何东晓, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, *, address=3天津大学智能与计算学部 天津 300350, bio={"img":"s+rhBDgLapnAhFkIkLmH3w==","content":"

1984年生. 博士,教授. 主要研究方向为复杂网络分析、图数据挖掘、图神经网络

"}, bioImg=s+rhBDgLapnAhFkIkLmH3w==, bioContent=

1984年生. 博士,教授. 主要研究方向为复杂网络分析、图数据挖掘、图神经网络

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288491938538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=3, ext=[AuthorCompanyExt(id=1329457288500327147, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Intelligence and Computing, Tianjin University, Tianjin 300350), AuthorCompanyExt(id=1329457288512910060, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3天津大学智能与计算学部 天津 300350)])]), Author(id=1329457293034369801, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=4, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457293135033100, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457293034369801, language=EN, stringName=Dong Liu, firstName=Dong, middleName=null, lastName=Liu, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=4, address=4College of Computer and Information Engineering, Henan Normal University, Xinxiang, Henan 453007, bio={"content":"

born in 1976. PhD, professor. His main research interests include educational data mining and complex network analysis

"}, bioImg=null, bioContent=

born in 1976. PhD, professor. His main research interests include educational data mining and complex network analysis

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457293197947662, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457293034369801, language=CN, stringName=刘栋, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=4, address=4河南师范大学计算机与信息工程学院 河南新乡 453007, bio={"img":"c+9oh/J1LNC/PRFDEP3NiA==","content":"

1976年生. 博士,教授. 主要研究方向为教育数据挖掘、复杂网络分析

"}, bioImg=c+9oh/J1LNC/PRFDEP3NiA==, bioContent=

1976年生. 博士,教授. 主要研究方向为教育数据挖掘、复杂网络分析

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288592601838, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=4, ext=[AuthorCompanyExt(id=1329457288596796143, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288592601838, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=4College of Computer and Information Engineering, Henan Normal University, Xinxiang, Henan 453007), AuthorCompanyExt(id=1329457288605184752, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288592601838, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=4河南师范大学计算机与信息工程学院 河南新乡 453007)])])], keywords=[Keyword(id=1329457293306999567, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, orderNo=1, keyword=graph machine learning), Keyword(id=1329457293382497040, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, orderNo=2, keyword=graph neural networks), Keyword(id=1329457293441217297, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, orderNo=3, keyword=neural ordinary differential equations), Keyword(id=1329457293516714770, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, orderNo=4, keyword=taxonomy), Keyword(id=1329457295131521812, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, orderNo=5, keyword=survey), Keyword(id=1329457295194436374, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, orderNo=1, keyword=图机器学习), Keyword(id=1329457295282516759, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, orderNo=2, keyword=图神经网络), Keyword(id=1329457295341237016, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, orderNo=3, keyword=神经常微分方程), Keyword(id=1329457295395762969, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, orderNo=4, keyword=分类体系), Keyword(id=1329457295475454746, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, orderNo=5, keyword=综述)], refs=[Reference(id=1329457297652298556, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=1, rfOrder=0, authorNames=null, journalName=null, refType=null, unstructuredReference=徐冰冰,岑科廷,黄俊杰,等. 图卷积神经网络综述[J]. 计算机学报,2020,43(5):755−780, articleTitle=null, refAbstract=null), Reference(id=1329457297719407421, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=1, rfOrder=1, authorNames=null, journalName=null, refType=null, unstructuredReference=Xu Bingbing, Cen Keting, Huang Junjie, et al. A survey on graph convolutional network[J]. Chinese Journal of Computers, 2020, 43(5): 755−780 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329457297782321982, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=2, rfOrder=2, authorNames=null, journalName=null, refType=null, unstructuredReference=马帅,刘建伟,左信. 图神经网络综述[J]. 计算机研究与发展,2022,59(1):47−80, articleTitle=null, refAbstract=null), Reference(id=1329457297857819456, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=2, rfOrder=3, authorNames=null, journalName=null, refType=null, unstructuredReference=Ma Shuai, Liu Jianwei, Zuo Xin. Survey on graph neural network[J]. Journal of Computer Research and Development, 2022, 59(1): 47−80 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329457297929122625, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=3, rfOrder=4, authorNames=null, journalName=null, refType=null, unstructuredReference=Wu Zonghan, Pan Shirui, Chen Fengwen, et al. A comprehensive survey on graph neural networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2020, 32(1): 4−24, articleTitle=null, refAbstract=null), Reference(id=1329457297983648578, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=4, rfOrder=5, authorNames=null, journalName=null, refType=null, unstructuredReference=Liu Juncheng, Hooi B, Kawaguchi K, et al. Scalable and effective implicit graph neural networks on large graphs[C/OL]//Proc of the 12th Int Conf on Learning Representations. 2024[2024-05-16].https://openreview.net/forum?id=QcMdPYBwTu, articleTitle=null, refAbstract=null), Reference(id=1329457299594261315, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=5, rfOrder=6, authorNames=null, journalName=null, refType=null, unstructuredReference=Xu Keyulu, Hu Weihua, Leskovec J, et al. How powerful are graph neural networks?[C/OL]//Proc of the 7th Int Conf on Learning Representations. 2019[2024-05-16].https://openreview.net/forum?id=ryGs6iA5Km, articleTitle=null, refAbstract=null), Reference(id=1329457299690730309, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=6, rfOrder=7, authorNames=null, journalName=null, refType=null, unstructuredReference=Gasteiger J, Bojchevski A, Günnemann S. Predict then propagate: Graph neural networks meet personalized pagerank[C/OL]//Proc of the 7th Int Conf on Learning Representations. 2019[2024-05-16].https://openreview.net/forum?id=H1gL-2A9Ym, articleTitle=null, refAbstract=null), Reference(id=1329457299770422086, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=7, rfOrder=8, authorNames=null, journalName=null, refType=null, unstructuredReference=Hajiramezanali E, Hasanzadeh A, Narayanan K, et al. Variational graph recurrent neural networks[C]//Proc of the 33rd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2019, 32: 10701−10711, articleTitle=null, refAbstract=null), Reference(id=1329457299845919559, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=8, rfOrder=9, authorNames=null, journalName=null, refType=null, unstructuredReference=Xu Da, Ruan Chuanwei, Korpeoglu E, et al. Inductive representation learning on temporal graphs[C/OL]//Proc of the 8th Int Conf on Learning Representations. 2020[2024-05-16].https://openreview.net/forum?id=rJeW1yHYwH, articleTitle=null, refAbstract=null), Reference(id=1329457299925611337, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=9, rfOrder=10, authorNames=null, journalName=null, refType=null, unstructuredReference=Wu Yongji, Lian Defu, Xu Yiheng, et al. Graph convolutional networks with Markov random field reasoning for social spammer detection[C]//Proc of the 34th AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2020, 34(1): 1054−1061, articleTitle=null, refAbstract=null), Reference(id=1329457300017886026, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=10, rfOrder=11, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Si, Tong Hanghang, Xu Jiejun, et al. Graph convolutional networks: A comprehensive review[J]. Computational Social Networks, 2019, 6(1): 1−23, articleTitle=null, refAbstract=null), Reference(id=1329457300101772107, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=11, rfOrder=12, authorNames=null, journalName=null, refType=null, unstructuredReference=Ying Rex, He Ruining, Chen Kaifeng, et al. Graph convolutional neural networks for web-scale recommender systems[C]//Proc of the 24th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2018: 974−983, articleTitle=null, refAbstract=null), Reference(id=1329457300223406925, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=12, rfOrder=13, authorNames=null, journalName=null, refType=null, unstructuredReference=王磊,熊于宁,李云鹏,等. 一种基于增强图卷积神经网络的协同推荐模型[J]. 计算机研究与发展,2021,58(9):1987−1996, articleTitle=null, refAbstract=null), Reference(id=1329457300294710094, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=12, rfOrder=14, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Lei, Xiong Yuning, Li Yunpeng, el al. A collaborative recommendation model based on enhanced graph convolutional neural network[J]. Journal of Computer Research and Development, 2021, 58(9): 1987−1996 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329457300353430351, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=13, rfOrder=15, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Shoujin, Hu Liang, Wang Yan, et al. Graph learning based recommender systems: A review[C]//Proc of the 30th Int Joint Conf on Artificial Intelligence. San Francisco, CA: IJCAI, 2021: 4644−4652, articleTitle=null, refAbstract=null), Reference(id=1329457300420539216, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=14, rfOrder=16, authorNames=null, journalName=null, refType=null, unstructuredReference=Fout A, Byrd J, Shariat B, et al. Protein interface prediction using graph convolutional networks[C]//Proc of the 31st Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2017, 30: 6533−6542, articleTitle=null, refAbstract=null), Reference(id=1329457300483453778, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=15, rfOrder=17, authorNames=null, journalName=null, refType=null, unstructuredReference=Borgwardt K M, Ong C S, Schönauer S, et al. Protein function prediction via graph kernels[J]. Bioinformatics, 2005, 21(suppl_1): i47−i56, articleTitle=null, refAbstract=null), Reference(id=1329457300550562643, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=16, rfOrder=18, authorNames=null, journalName=null, refType=null, unstructuredReference=Guo Shengnan, Lin Youfang, Feng Ning, et al. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting[C]//Proc of the 33rd AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2019: 922−929, articleTitle=null, refAbstract=null), Reference(id=1329457300617671508, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=17, rfOrder=19, authorNames=null, journalName=null, refType=null, unstructuredReference=Yu Bing, Yin Haoteng, Zhu Zhanxing. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting[C]//Proc of the 28th Int Joint Conf on Artificial Intelligence. San Francisco, CA: IJCAI, 2018: 3634−3640, articleTitle=null, refAbstract=null), Reference(id=1329457300680586069, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=18, rfOrder=20, authorNames=null, journalName=null, refType=null, unstructuredReference=杜圣东,李天瑞,杨燕,等. 一种基于序列到序列时空注意力学习的交通流预测模型[J]. 计算机研究与发展,2020,57(8):1715−1728, articleTitle=null, refAbstract=null), Reference(id=1329457300747694935, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=18, rfOrder=21, authorNames=null, journalName=null, refType=null, unstructuredReference=Du Shengdong, Li Tianrui, Yang Yan, et al. A sequence-to-sequence spatial-temporal attention learning model for urban traffic flow prediction[J]. Journal of Computer Research and Development, 2020, 57(8): 1715−1728 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329457300802220887, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=19, rfOrder=22, authorNames=null, journalName=null, refType=null, unstructuredReference=Bodnar C, Di Giovanni F, Chamberlain B, et al. Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in GNNS[C]//Proc of the 36th Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2022, 35: 18527−18541, articleTitle=null, refAbstract=null), Reference(id=1329457300856746840, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=20, rfOrder=23, authorNames=null, journalName=null, refType=null, unstructuredReference=Rusch T K, Bronstein M M, Mishra S. A survey on oversmoothing in graph neural networks[J]. arXiv preprint, arXiv: 2303.10993, 2023, articleTitle=null, refAbstract=null), Reference(id=1329457300923855705, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=21, rfOrder=24, authorNames=null, journalName=null, refType=null, unstructuredReference=Rong Yu, Huang Wenbing, Xu Tingyang, et al. DropEdge: Towards deep graph convolutional networks on node classification[C/OL]//Proc of the 8th Int Conf on Learning Representations. 2020 [2024-05-16].https://openreview.net/forum?id=Hkx1qkrKPr, articleTitle=null, refAbstract=null), Reference(id=1329457300978381658, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=22, rfOrder=25, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhao L, Akoglu L. PairNorm: Tackling oversmoothing in GNN[C]//Proc of the 8th Int Conf on Learning Representations. 2020[2024-05-16].https://openreview.net/forum?id=rkecl1rtwB, articleTitle=null, refAbstract=null), Reference(id=1329457301041296219, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=23, rfOrder=26, authorNames=null, journalName=null, refType=null, unstructuredReference=Xu Keyulu, Li Chengtao, Tian Yonglong, et al. Representation learning on graphs with jumping knowledge networks[C]//Proc of the 35th Int Conf on Machine Learning. New York: PMLR, 2018: 5453−5462, articleTitle=null, refAbstract=null), Reference(id=1329457301091627869, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=24, rfOrder=27, authorNames=null, journalName=null, refType=null, unstructuredReference=Barzel B, Liu Yangyu, Barabási A L. Constructing minimal models for complex system dynamics[J]. Nature Communication, 2015, 6(1): 7186, articleTitle=null, refAbstract=null), Reference(id=1329457301146153822, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=25, rfOrder=28, authorNames=null, journalName=null, refType=null, unstructuredReference=Alvarez-Rodriguez U, Battiston F, de Arruda G F, et al. Evolutionary dynamics of higher-order interactions in social networks[J]. Nature Human Behaviour, 2021, 5(5): 586−595, articleTitle=null, refAbstract=null), Reference(id=1329457301204874079, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=26, rfOrder=29, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Yanfu, Gao Shangqian, Pei Jian, et al. Improving social network embedding via new second-order continuous graph neural networks[C]//Proc of the 28th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2022: 2515−2523, articleTitle=null, refAbstract=null), Reference(id=1329457301271982944, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=27, rfOrder=30, authorNames=null, journalName=null, refType=null, unstructuredReference=Hou Jinlin, Guo Xuan, Liu Jiye, et al. Structure-enhanced graph neural ODE network for temporal link prediction[C]//Proc of the 33rd Int Conf on Artificial Neural Networks. Berlin: Springer, 2023: 563−575, articleTitle=null, refAbstract=null), Reference(id=1329457301339091809, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=28, rfOrder=31, authorNames=null, journalName=null, refType=null, unstructuredReference=Fey M, Lenssen J E, Weichert F, et al. Gnnautoscale: Scalable and expressive graph neural networks via historical embeddings[C]//Proc of the 38th Int Confon Machine Learning. New York: PMLR, 2021: 3294−3304, articleTitle=null, refAbstract=null), Reference(id=1329457301406200674, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=29, rfOrder=32, authorNames=null, journalName=null, refType=null, unstructuredReference=Chen R T Q, Rubanova Y, Bettencourt J, et al. Neural ordinary differential equations[C]//Proc of the 32nd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2018, 31: 6572−6583, articleTitle=null, refAbstract=null), Reference(id=1329457301481698148, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=30, rfOrder=33, authorNames=null, journalName=null, refType=null, unstructuredReference=Xhonneux L P, Qu Meng, Tang Jian. Continuous graph neural networks[C]//Proc of the 37th Int Conf on Machine Learning. New York: PMLR, 2020: 10432−10441, articleTitle=null, refAbstract=null), Reference(id=1329457301553001317, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=31, rfOrder=34, authorNames=null, journalName=null, refType=null, unstructuredReference=Zang Chengxi, Wang Fei. Neural dynamics on complex networks[C]//Proc of the 26th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2020: 892−902, articleTitle=null, refAbstract=null), Reference(id=1329457301624304486, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=32, rfOrder=35, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Zihui, Yang Peizhen, Fan Xiaoliang, et al. ConTIG: Continuous representation learning on temporal interaction graphs[J]. Neural Networks, 2024, 172: 106151, articleTitle=null, refAbstract=null), Reference(id=1329457301703996264, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=33, rfOrder=36, authorNames=null, journalName=null, refType=null, unstructuredReference=Veličković P, Cucurull G, Casanova A, et al. Graph attention networks[C/OL]//Proc of the 6th Int Conf on Learning Representations. 2018[2024-05-16].https://openreview.net/forum?id=rJXMpikCZ, articleTitle=null, refAbstract=null), Reference(id=1329457301775299433, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=34, rfOrder=37, authorNames=null, journalName=null, refType=null, unstructuredReference=Rossi E, Chamberlain B, Frasca F, et al. Temporal graph networks for deep learning on dynamic graphs[C]//Proc of the ICML 2020 Workshop on Graph Representation Learning and Beyond. New York: PMLR, 2020: 1−9, articleTitle=null, refAbstract=null), Reference(id=1329457301917905770, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=35, rfOrder=38, authorNames=null, journalName=null, refType=null, unstructuredReference=Ma Yao, Guo Ziyi, Ren Zhaocun, et al. Streaming graph neural networks[C]//Proc of the 43rd Int ACM SIGIR Conf on Research and Development in Information Retrieval. New York: ACM, 2020: 719−728, articleTitle=null, refAbstract=null), Reference(id=1329457301985014635, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=36, rfOrder=39, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhou Hongkuan, Zheng Da, Nisa I, et al. TGL: A general framework for temporal gnn training on billion-scale graphs[C]//Proc of the 48th Int Conf on Very Large Databases. New York: VLDB Endowment, 2022, 15(8): 1572−1580, articleTitle=null, refAbstract=null), Reference(id=1329457302043734892, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=37, rfOrder=40, authorNames=null, journalName=null, refType=null, unstructuredReference=Pareja A, Domeniconi G, Chen Jie, et al. EvolveGCN: Evolving graph convolutional networks for dynamic graphs[C]//Proc of the 34th AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2020: 5363−5370, articleTitle=null, refAbstract=null), Reference(id=1329457302106649453, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=38, rfOrder=41, authorNames=null, journalName=null, refType=null, unstructuredReference=Taheri A, Gimpel K, Berger-Wolf T. Learning to represent the evolution of dynamic graphs with recurrent models[C]//Proc of Companion Proc of the 2019 World Wide Web Conf. New York: ACM, 2019: 301−307, articleTitle=null, refAbstract=null), Reference(id=1329457302169564014, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=39, rfOrder=42, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Yanbang, Li Pan, Bai Chongyang, et al. Generic representation learning for dynamic social interaction[C]//Proc of the 26th ACM SIGKDD Conf on Knowledge Discovery and Data Mining Workshop. New York: ACM, 2020: 1−9, articleTitle=null, refAbstract=null), Reference(id=1329457302228284272, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=40, rfOrder=43, authorNames=null, journalName=null, refType=null, unstructuredReference=You Jiaxuan, Du Tianyu, Leskovec J. ROLAND: Graph learning framework for dynamic graphs[C]//Proc of the 28th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2022: 2358−2366, articleTitle=null, refAbstract=null), Reference(id=1329457302303781745, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=41, rfOrder=44, authorNames=null, journalName=null, refType=null, unstructuredReference=Pontryagin L S. Mathematical Theory of Optimal Processes[M]. London: Routledge, 2018, articleTitle=null, refAbstract=null), Reference(id=1329457302370890610, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=42, rfOrder=45, authorNames=null, journalName=null, refType=null, unstructuredReference=Rubanova Y, Chen R T Q, Duvenaud D K. Latent ordinary differential equations for irregularly-sampled time series[C]//Proc of the 33rd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2019, 32: 5320−5330, articleTitle=null, refAbstract=null), Reference(id=1329457302429610867, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=43, rfOrder=46, authorNames=null, journalName=null, refType=null, unstructuredReference=Dupont E, Doucet A, Teh Y W. Augmented neural ODEs[C]//Proc of the 33rd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2019, 32: 3140−3150, articleTitle=null, refAbstract=null), Reference(id=1329457304044417909, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=44, rfOrder=47, authorNames=null, journalName=null, refType=null, unstructuredReference=Biswas B N, Chatterjee S, Mukherjee S P, et al. A discussion on Euler method: A review[J]. Electronic Journal of Mathematical Analysis and Applications, 2013, 1(2): 2090−2792, articleTitle=null, refAbstract=null), Reference(id=1329457305998963574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=45, rfOrder=48, authorNames=null, journalName=null, refType=null, unstructuredReference=Dormand J R, Prince P J. A family of embedded Runge-Kutta formulae[J]. Journal of Computational and Applied Mathematics, 1980, 6(1): 19−26, articleTitle=null, refAbstract=null), Reference(id=1329457306095432568, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=46, rfOrder=49, authorNames=null, journalName=null, refType=null, unstructuredReference=Fang Zheng, Long Qingqing, Song Guojie, et al. Spatial-temporal graph ode networks for traffic flow forecasting[C]//Proc of the 27th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2021: 364−373, articleTitle=null, refAbstract=null), Reference(id=1329457306154152825, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=47, rfOrder=50, authorNames=null, journalName=null, refType=null, unstructuredReference=Jin Ming, Zheng Yu, Li Yuanfang, et al. Multivariate time series forecasting with dynamic graph neural ODES[J]. IEEE Transactions on Knowledge and Data Engineering, 2023, 35(9): 9168−80, articleTitle=null, refAbstract=null), Reference(id=1329457306221261690, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=48, rfOrder=51, authorNames=null, journalName=null, refType=null, unstructuredReference=Huang Zijie, Sun Yizhou, Wang Wei. Coupled graph ODE for learning interacting system dynamics[C]//Proc of the 27th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2021: 705−715, articleTitle=null, refAbstract=null), Reference(id=1329457306288370555, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=49, rfOrder=52, authorNames=null, journalName=null, refType=null, unstructuredReference=Choi J, Jeon J, Park N. LT-OCF: Learnable-time ode-based collaborative filtering[C]//Proc of the 30th ACM Int Conf on Information and Knowledge Management. New York, ACM, 2021: 251−260, articleTitle=null, refAbstract=null), Reference(id=1329457306347090812, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=50, rfOrder=53, authorNames=null, journalName=null, refType=null, unstructuredReference=Jiang Song, Huang Zijie, Luo Xiao, et al. CF-GODE: Continuous-time causal inference for multi-agent dynamical systems[C]//Proc of the 29th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2023: 997−1009, articleTitle=null, refAbstract=null), Reference(id=1329457306422588286, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=51, rfOrder=54, authorNames=null, journalName=null, refType=null, unstructuredReference=Rusch T K, Chamberlain B, Rowbottom J, et al. Graph-coupled oscillator networks[C]//Proc of the 39th Int Conf on Machine Learning. New York: PMLR, 2022: 18888−18909, articleTitle=null, refAbstract=null), Reference(id=1329457306498085759, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=52, rfOrder=55, authorNames=null, journalName=null, refType=null, unstructuredReference=Huang Zijie, Zhao Wanjie, Gao Jingdong, et al. TANGO: Time-reversal latent GraphODE for multi-agent dynamical systems[C/OL]//Proc of NeurIPS 2023 Workshop on the Symbiosis of Deep Learning and Differential Equations. 2023[2024-05-16]. https://openreview.net/forum?id=8AYetbupoZ, articleTitle=null, refAbstract=null), Reference(id=1329457306573583232, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=53, rfOrder=56, authorNames=null, journalName=null, refType=null, unstructuredReference=Liu Yang, Cheng Jiashun, Zhao Haihong, et al. Physics-inspired neural graph ODE for long-term dynamical simulation[J]. arXiv preprint, arXiv: 2308.13212, 2023, articleTitle=null, refAbstract=null), Reference(id=1329457306649080705, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=54, rfOrder=57, authorNames=null, journalName=null, refType=null, unstructuredReference=Schober M, Särkkä S, Hennig P. A probabilistic model for the numerical solution of initial value problems[J]. Statistics and Computing, 2019, 29(1): 99−122, articleTitle=null, refAbstract=null), Reference(id=1329457306728772482, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=55, rfOrder=58, authorNames=null, journalName=null, refType=null, unstructuredReference=Poli M, Massaroli S, Park J, et al. Graph neural ordinary differential equations [J]. arXiv preprint, arXiv: 1911.07532, 2019, articleTitle=null, refAbstract=null), Reference(id=1329457306800075652, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=56, rfOrder=59, authorNames=null, journalName=null, refType=null, unstructuredReference=Huang Zijie, Sun Yizhou, Wang Wei. Learning continuous system dynamics from irregularly-sampled partial observations[C]//Proc of the 34th Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2020, 33: 16177−16187, articleTitle=null, refAbstract=null), Reference(id=1329457306875573125, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=57, rfOrder=60, authorNames=null, journalName=null, refType=null, unstructuredReference=Huang Zijie, Sun Yizhou, Wang Wei. Generalizing graph ODE for learning complex system dynamics across environments[C]//Proc of the 29th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2023: 798−809, articleTitle=null, refAbstract=null), Reference(id=1329457306938487686, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=58, rfOrder=61, authorNames=null, journalName=null, refType=null, unstructuredReference=Hwang J, Choi J, Choi H, et al. Climate modeling with neural diffusion equations[C]//Proc of 2021 IEEE Int Conf on Data Mining. Piscataway, NJ: IEEE, 2021: 230−239, articleTitle=null, refAbstract=null), Reference(id=1329457308574266247, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=59, rfOrder=62, authorNames=null, journalName=null, refType=null, unstructuredReference=Chen Yibi, Qin Yunchuan, Li Kenli, et al. Adaptive spatial-temporal graph convolution networks for collaborative local-global learning in traffic prediction[J]. IEEE Transactions on Vehicular Technology, 2023, 72(10): 12653−12663, articleTitle=null, refAbstract=null), Reference(id=1329457308658152329, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=60, rfOrder=63, authorNames=null, journalName=null, refType=null, unstructuredReference=Guo Jiayan, Zhang Peiyan, Li Chaozhuo, et al. Evolutionary preference learning via graph nested GRU ODE for session-based recommendation[C]//Proc of the 31st ACM Int Conf on Information and Knowledge Management. New York: ACM, 2022: 624−634, articleTitle=null, refAbstract=null), Reference(id=1329457308750427018, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=61, rfOrder=64, authorNames=null, journalName=null, refType=null, unstructuredReference=Luo Xiao, Yuan Jingyang, Huang Zijie, et al. HOPE: High-order graph ODE for modeling interacting dynamics[C]//Proc of the 40th Int Conf on Machine Learning. New York: PMLR, 2023: 23124−23139, articleTitle=null, refAbstract=null), Reference(id=1329457308851090316, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=62, rfOrder=65, authorNames=null, journalName=null, refType=null, unstructuredReference=Jin Ming, Li Yuanfang, Pan Shirui. Neural temporal walks: Motif-aware representation learning on continuous-time dynamic graphs[C]//Proc of the 36th Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2022, 35: 19874−19886, articleTitle=null, refAbstract=null), Reference(id=1329457308934976397, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=63, rfOrder=66, authorNames=null, journalName=null, refType=null, unstructuredReference=Qin Yifang, Ju Wei, Wu Hongjun, et al. Learning graph node for continuous-time sequential recommendation [J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(7): 1−14, articleTitle=null, refAbstract=null), Reference(id=1329457309006279566, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=64, rfOrder=67, authorNames=null, journalName=null, refType=null, unstructuredReference=Gao Zhihan, Wang Hao, Wang Yuyang, et al. Probabilistic continuous-time whole-graph forecasting[C]//Proc of the 8th SIGKDD Int Workshop on Mining and Learning from Time Series. New York: ACM, 2022: 1−14, articleTitle=null, refAbstract=null), Reference(id=1329457309081777039, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=65, rfOrder=68, authorNames=null, journalName=null, refType=null, unstructuredReference=Kipf T N, Welling M. Semi-supervised classification with graph convolutional networks[C/OL]//Proc of the 5th Int Conf on Learning Representations. 2017[2024-05-16]. https://openreview.net/forum?id=SJU4ayYgl, articleTitle=null, refAbstract=null), Reference(id=1329457309157274513, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=66, rfOrder=69, authorNames=null, journalName=null, refType=null, unstructuredReference=Skarding J, Gabrys B, Musial K. Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey[J]. IEEE Access, 2021, 9: 79143−79168, articleTitle=null, refAbstract=null), Reference(id=1329457309274715026, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=67, rfOrder=70, authorNames=null, journalName=null, refType=null, unstructuredReference=Sankar A, Wu Yanhong, Gou Liang, et al. DySAT: Deep neural representation learning on dynamic graphs via self-attention networks[C]//Proc of the 13th Int Conf on Web Search and Data Mining. New York: ACM, 2020: 519−527, articleTitle=null, refAbstract=null), Reference(id=1329457309358601108, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=68, rfOrder=71, authorNames=null, journalName=null, refType=null, unstructuredReference=Min Shengjie, Gao Zhan, Peng Jing, et al. STGSN—a spatial temporal graph neural network framework for time-evolving social networks[J]. Knowledge-Based Systems, 2021, 214: 106746, articleTitle=null, refAbstract=null), Reference(id=1329457309480235925, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=69, rfOrder=72, authorNames=null, journalName=null, refType=null, unstructuredReference=Gao Jianfei, Ribeiro B. On the equivalence between temporal and static equivariant graph representations[C]//Proc of the 39th Int Conf on Machine Learning. New York: PMLR, 2022: 7052−7076, articleTitle=null, refAbstract=null), Reference(id=1329457309551539094, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=70, rfOrder=73, authorNames=null, journalName=null, refType=null, unstructuredReference=Yao Liuyi, Chu Zhixuan, Li Sheng, et al. A survey on causal inference[J]. ACM Transactions on Knowledge Discovery from Data, 2021, 15(5): 1−46, articleTitle=null, refAbstract=null)], funds=null, companyList=[AuthorCompany(id=1329457288340943588, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=1, ext=[AuthorCompanyExt(id=1329457288349332197, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288340943588, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Cyberspace, Hangzhou Dianzi University, Hangzhou 310018), AuthorCompanyExt(id=1329457288357720806, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288340943588, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1杭州电子科技大学网络空间安全学院 杭州 310018)]), AuthorCompany(id=1329457288408052455, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=2, ext=[AuthorCompanyExt(id=1329457288412246760, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288408052455, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2ZhuoYue Honors College, Hangzhou Dianzi University, Hangzhou 310018), AuthorCompanyExt(id=1329457288420635369, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288408052455, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2杭州电子科技大学卓越学院 杭州 310018)]), AuthorCompany(id=1329457288491938538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=3, ext=[AuthorCompanyExt(id=1329457288500327147, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Intelligence and Computing, Tianjin University, Tianjin 300350), AuthorCompanyExt(id=1329457288512910060, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3天津大学智能与计算学部 天津 300350)]), AuthorCompany(id=1329457288592601838, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=4, ext=[AuthorCompanyExt(id=1329457288596796143, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288592601838, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=4College of Computer and Information Engineering, Henan Normal University, Xinxiang, Henan 453007), AuthorCompanyExt(id=1329457288605184752, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288592601838, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=4河南师范大学计算机与信息工程学院 河南新乡 453007)])], figs=[ArticleFig(id=1329457295605478172, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Figure 1, caption=Taxonomy of graph neural ODE methods, figureFileSmall=s5Hcpavyg7xu9+Fmlb0e8g==, figureFileBig=x9Yp13Vs2ygQxslfTGMNnA==, tableContent=null), ArticleFig(id=1329457295668392733, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=图1, caption=图神经ODE方法分类, figureFileSmall=s5Hcpavyg7xu9+Fmlb0e8g==, figureFileBig=x9Yp13Vs2ygQxslfTGMNnA==, tableContent=null), ArticleFig(id=1329457295878107935, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Figure 2, caption=Comparison of node classification performance and representation differences between some GNN and graph neural ODE, figureFileSmall=emMjr7dmotKZjbgLy6HfJw==, figureFileBig=fHg2rvbjE/OBe1hyS3YpVw==, tableContent=null), ArticleFig(id=1329457295936828192, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=图2, caption=部分GNN和图神经ODE节点分类效果和表示差异对比, figureFileSmall=emMjr7dmotKZjbgLy6HfJw==, figureFileBig=fHg2rvbjE/OBe1hyS3YpVw==, tableContent=null), ArticleFig(id=1329457295999742753, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Figure 3, caption=Network dynamics fitting results of some graph recurrent neural network and NDCN, figureFileSmall=Zwsk4e2+50p+Oy4kFhjtJA==, figureFileBig=XGlAxVXZ8wf9FTsZd0AZhQ==, tableContent=null), ArticleFig(id=1329457296092017443, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=图3, caption=部分图循环神经网络和NDCN网络动力学拟合结果, figureFileSmall=Zwsk4e2+50p+Oy4kFhjtJA==, figureFileBig=XGlAxVXZ8wf9FTsZd0AZhQ==, tableContent=null), ArticleFig(id=1329457296171709220, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Figure 4, caption=The results of link predicted on snapshot missing networks by some temporal GNN and SEGODE, figureFileSmall=rpNHDhaGAB1lRxvjtat62A==, figureFileBig=7pQLVkci4pZO/IlFOqb1MA==, tableContent=null), ArticleFig(id=1329457296238818085, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=图4, caption=部分时序图神经网络和SEG-ODE在快照缺失网络上的链路预测结果, figureFileSmall=rpNHDhaGAB1lRxvjtat62A==, figureFileBig=7pQLVkci4pZO/IlFOqb1MA==, tableContent=null), ArticleFig(id=1329457296297538342, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Figure 5, caption=Overhead comparison of some GNN and graph neural ODE on Cora dataset, figureFileSmall=5XMa6kbnhc55ceTdiabsxg==, figureFileBig=zPKy02xK3yt4pA2cjJQDoA==, tableContent=null), ArticleFig(id=1329457296360452903, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=图5, caption=Cora数据集上部分GNN和图神经ODE开销对比, figureFileSmall=5XMa6kbnhc55ceTdiabsxg==, figureFileBig=zPKy02xK3yt4pA2cjJQDoA==, tableContent=null), ArticleFig(id=1329457296448533288, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 1, caption=

Main Notations and Their Definitions

, figureFileSmall=null, figureFileBig=null, tableContent=
符号定义
$ G=(\mathrm{ }\mathcal{V},\mathcal{E},{\boldsymbol X}) $静态图
$ {\mathcal{G}}_{\mathrm{D}}=\{{G}^{1},{G}^{2},… ,{G}^{T}\} $离散动态图
$ {\mathcal{G}}_{\mathrm{C}}=\left\{\left({u}_{i},{v}_{i},{t}_{i}\right)\right\} $连续动态图
$ v,e $节点,边
$ {\boldsymbol A} $邻接矩阵
$ {\boldsymbol X} $特征矩阵
$ {\boldsymbol D} $度矩阵
$ \mathcal{N}\left(v\right) $节点$ v $的邻居节点集合
$ \sigma \left(\cdot \right) $激活函数
$ {\boldsymbol \theta } $可学习参数
$ {{\boldsymbol h }}_{v} $节点$ v $的表示
$ {\boldsymbol H} $所有节点的表示矩阵
$ {{\boldsymbol b }}_{ij} $节点对$ ({v}_{i},{v}_{j}) $的表示
), ArticleFig(id=1329457296507253545, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表1, caption=

主要符号及其定义

, figureFileSmall=null, figureFileBig=null, tableContent=
符号定义
$ G=(\mathrm{ }\mathcal{V},\mathcal{E},{\boldsymbol X}) $静态图
$ {\mathcal{G}}_{\mathrm{D}}=\{{G}^{1},{G}^{2},… ,{G}^{T}\} $离散动态图
$ {\mathcal{G}}_{\mathrm{C}}=\left\{\left({u}_{i},{v}_{i},{t}_{i}\right)\right\} $连续动态图
$ v,e $节点,边
$ {\boldsymbol A} $邻接矩阵
$ {\boldsymbol X} $特征矩阵
$ {\boldsymbol D} $度矩阵
$ \mathcal{N}\left(v\right) $节点$ v $的邻居节点集合
$ \sigma \left(\cdot \right) $激活函数
$ {\boldsymbol \theta } $可学习参数
$ {{\boldsymbol h }}_{v} $节点$ v $的表示
$ {\boldsymbol H} $所有节点的表示矩阵
$ {{\boldsymbol b }}_{ij} $节点对$ ({v}_{i},{v}_{j}) $的表示
), ArticleFig(id=1329457296574362410, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 2, caption=

ODE Numerical Solution Methods

, figureFileSmall=null, figureFileBig=null, tableContent=
方法形式化描述方法分类模型实例
显式Euler[44]$ \begin{array}{c}{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\Delta t\cdot f\left({\boldsymbol h }\left(t\right),t\right)\end{array} $显式、单步STGODE[46], MTGODE[47] CG-ODE[48], LT-OCF[49], NDCN[31], CF-GODE[50], GraphCON[51], TANGO[52], PINGO[53]
隐式Euler[44]$ \begin{array}{c}{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\Delta t\cdot f\left({\boldsymbol h }\left(t+\Delta t\right),t+\Delta t\right)\end{array} $隐式、单步
RK4[54]$ \begin{array}{l}{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{6}\left({k}_{1}+2{k}_{2}+2{k}_{3}+{k}_{4}\right)\\ {k}_{1}=f\left({\boldsymbol h }\left(t\right),t\right)\\ {k}_{2}=f\left({\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{2}{k}_{1},t+\dfrac{\Delta t}{2}\right)\\ {k}_{3}=f\left({\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{2}{k}_{2},t+\dfrac{\Delta t}{2}\right)\\ {k}_{4}=f\left({\boldsymbol h }\left(t\right)+\Delta t{k}_{3},t+\Delta t\right)\end{array} $显式、单步MTGODE[47], SEGODE[27], CG-ODE[48], NDCN[31],
LT-OCF[49], GCDE[55], LG-ODE[56], GG-ODE[57]
Adams-Moulton[49]$ \begin{aligned} {\boldsymbol h }\left(t+\Delta t\right)=&{\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{24}(9f\left({\boldsymbol h }\left(t+\Delta t\right),t+\Delta t\right)+19f\left({\boldsymbol h }\left(t\right),t\right)-\\&5f\left({\boldsymbol h }\left(t-\Delta t\right),t-\Delta t\right)+f\left({\boldsymbol h }\left(t-2\Delta t\right),t-2\Delta t\right))\end{aligned} $隐式、多步LT-OCF[49]
DOPRI[45]$ \begin{aligned}&{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\Delta t\displaystyle\sum _{i=1}^{7}{b}_{i}{k}_{i}\\&{k}_{1}=f\left({\boldsymbol h }\left(t\right),t\right)\\&{k}_{2}=f\left({\boldsymbol h }\left(t\right)+{a}_{21}\Delta t{k}_{1},t+{c}_{2}\Delta t\right)\\&{k}_{3}=f\left({\boldsymbol h }\left(t\right)+{a}_{31}\Delta t{k}_{1}+{a}_{32}\Delta t{k}_{2},t+{c}_{3}\Delta t\right)\\&\vdots \\&{k}_{7}=f\left({\boldsymbol h }\left(t\right)+{a}_{71}\Delta t{k}_{1}+{a}_{72}\Delta t{k}_{2}+… +{a}_{76}\Delta t{k}_{6},t+{c}_{7}\Delta t\right)\end{aligned} $显式、单步NDCN[31], LT-OCF[49], GCDE[55], NDE[58]
), ArticleFig(id=1329457296641471275, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表2, caption=

ODE数值求解方法

, figureFileSmall=null, figureFileBig=null, tableContent=
方法形式化描述方法分类模型实例
显式Euler[44]$ \begin{array}{c}{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\Delta t\cdot f\left({\boldsymbol h }\left(t\right),t\right)\end{array} $显式、单步STGODE[46], MTGODE[47] CG-ODE[48], LT-OCF[49], NDCN[31], CF-GODE[50], GraphCON[51], TANGO[52], PINGO[53]
隐式Euler[44]$ \begin{array}{c}{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\Delta t\cdot f\left({\boldsymbol h }\left(t+\Delta t\right),t+\Delta t\right)\end{array} $隐式、单步
RK4[54]$ \begin{array}{l}{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{6}\left({k}_{1}+2{k}_{2}+2{k}_{3}+{k}_{4}\right)\\ {k}_{1}=f\left({\boldsymbol h }\left(t\right),t\right)\\ {k}_{2}=f\left({\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{2}{k}_{1},t+\dfrac{\Delta t}{2}\right)\\ {k}_{3}=f\left({\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{2}{k}_{2},t+\dfrac{\Delta t}{2}\right)\\ {k}_{4}=f\left({\boldsymbol h }\left(t\right)+\Delta t{k}_{3},t+\Delta t\right)\end{array} $显式、单步MTGODE[47], SEGODE[27], CG-ODE[48], NDCN[31],
LT-OCF[49], GCDE[55], LG-ODE[56], GG-ODE[57]
Adams-Moulton[49]$ \begin{aligned} {\boldsymbol h }\left(t+\Delta t\right)=&{\boldsymbol h }\left(t\right)+\dfrac{\Delta t}{24}(9f\left({\boldsymbol h }\left(t+\Delta t\right),t+\Delta t\right)+19f\left({\boldsymbol h }\left(t\right),t\right)-\\&5f\left({\boldsymbol h }\left(t-\Delta t\right),t-\Delta t\right)+f\left({\boldsymbol h }\left(t-2\Delta t\right),t-2\Delta t\right))\end{aligned} $隐式、多步LT-OCF[49]
DOPRI[45]$ \begin{aligned}&{\boldsymbol h }\left(t+\Delta t\right)={\boldsymbol h }\left(t\right)+\Delta t\displaystyle\sum _{i=1}^{7}{b}_{i}{k}_{i}\\&{k}_{1}=f\left({\boldsymbol h }\left(t\right),t\right)\\&{k}_{2}=f\left({\boldsymbol h }\left(t\right)+{a}_{21}\Delta t{k}_{1},t+{c}_{2}\Delta t\right)\\&{k}_{3}=f\left({\boldsymbol h }\left(t\right)+{a}_{31}\Delta t{k}_{1}+{a}_{32}\Delta t{k}_{2},t+{c}_{3}\Delta t\right)\\&\vdots \\&{k}_{7}=f\left({\boldsymbol h }\left(t\right)+{a}_{71}\Delta t{k}_{1}+{a}_{72}\Delta t{k}_{2}+… +{a}_{76}\Delta t{k}_{6},t+{c}_{7}\Delta t\right)\end{aligned} $显式、单步NDCN[31], LT-OCF[49], GCDE[55], NDE[58]
), ArticleFig(id=1329457296721163053, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 3, caption=

Static Graph Neural ODE

, figureFileSmall=null, figureFileBig=null, tableContent=
算法ODE函数ODE的机制创新ODE的作用
CGNN[30]$ \left\{\begin{aligned}&\dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=(\widehat{{\boldsymbol A}}-{\boldsymbol I}){\boldsymbol H}\left(\tau \right)+{\boldsymbol H}\left(0\right)\\ &\dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=(\widehat{{\boldsymbol A}}-{\boldsymbol I}){\boldsymbol H}\left(\tau \right)+{\boldsymbol H}\left(\tau \right)\left({\boldsymbol W}-{\boldsymbol I}\right)+{\boldsymbol H}\left(0\right)\end{aligned}\right. $提出考虑多特征信道交互的ODE编码器
SANODE[26]$ \dfrac{{d}^{2}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}{\tau }^{2}}+\alpha \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau }={f}_{\mathrm{F}}\left({{\boldsymbol h }}_{v}\left(\tau \right)\right)+{f}_{\mathrm{g}}\left({{\boldsymbol h }}_{v}\left(\tau \right),{{\boldsymbol h }}_{\mathcal{N}\left(v\right)}\left(\tau \right)\right) $基于消息传递的二阶ODE
GraphCON[51]$ \dfrac{{d}^{2}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}{\tau }^{2}}+\alpha \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau }=\sigma \left(f\left({\boldsymbol X},\tau ;{\boldsymbol \theta }\right)\right)-\gamma {\boldsymbol X} $基于物理系统的二阶ODE
NDE[58]$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=-{\boldsymbol K}\odot {\boldsymbol L}{\boldsymbol H}\left(\tau \right)+f({\boldsymbol H}\left(\tau \right),\tau ;{\boldsymbol \theta }) $引入了离散扩散算子
GCDE[55]$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }={G}{C}{N}({\boldsymbol H}\left(\tau \right),{\boldsymbol \varTheta }) $基于GCN的ODE
), ArticleFig(id=1329457296784077614, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表3, caption=

静态图神经ODE

, figureFileSmall=null, figureFileBig=null, tableContent=
算法ODE函数ODE的机制创新ODE的作用
CGNN[30]$ \left\{\begin{aligned}&\dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=(\widehat{{\boldsymbol A}}-{\boldsymbol I}){\boldsymbol H}\left(\tau \right)+{\boldsymbol H}\left(0\right)\\ &\dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=(\widehat{{\boldsymbol A}}-{\boldsymbol I}){\boldsymbol H}\left(\tau \right)+{\boldsymbol H}\left(\tau \right)\left({\boldsymbol W}-{\boldsymbol I}\right)+{\boldsymbol H}\left(0\right)\end{aligned}\right. $提出考虑多特征信道交互的ODE编码器
SANODE[26]$ \dfrac{{d}^{2}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}{\tau }^{2}}+\alpha \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau }={f}_{\mathrm{F}}\left({{\boldsymbol h }}_{v}\left(\tau \right)\right)+{f}_{\mathrm{g}}\left({{\boldsymbol h }}_{v}\left(\tau \right),{{\boldsymbol h }}_{\mathcal{N}\left(v\right)}\left(\tau \right)\right) $基于消息传递的二阶ODE
GraphCON[51]$ \dfrac{{d}^{2}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}{\tau }^{2}}+\alpha \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau }=\sigma \left(f\left({\boldsymbol X},\tau ;{\boldsymbol \theta }\right)\right)-\gamma {\boldsymbol X} $基于物理系统的二阶ODE
NDE[58]$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=-{\boldsymbol K}\odot {\boldsymbol L}{\boldsymbol H}\left(\tau \right)+f({\boldsymbol H}\left(\tau \right),\tau ;{\boldsymbol \theta }) $引入了离散扩散算子
GCDE[55]$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }={G}{C}{N}({\boldsymbol H}\left(\tau \right),{\boldsymbol \varTheta }) $基于GCN的ODE
), ArticleFig(id=1329457296842797871, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 4, caption=

Discrete Dynamic Graph Neural ODE

, figureFileSmall=null, figureFileBig=null, tableContent=
算法ODE函数$ {\widehat{\boldsymbol H}}^{t} $ODE作用
NDCN[31]$ \dfrac{{\rm d}{\boldsymbol H'^{t}}}{{\rm d}t}=f\left({\boldsymbol H'},G,{{\boldsymbol W}}_{\mathrm{h}},t\right) $$ {f}_{\mathrm{E}}\left({{\boldsymbol H}}^{\mathit{t}},{{\boldsymbol W}}_{\mathrm{E}}\right) $推断器
STG-ODE[46]$ \dfrac{\mathrm{d}\mathcal{H}\left(\tau \right)}{\mathrm{d}\tau }=\mathcal{H}\left(\tau \right){\times }_{1}\left({\widehat{\boldsymbol A}}-{\boldsymbol I}\right)+\mathcal{H}\left(\tau \right){\times }_{2}\left({\boldsymbol U}-{\boldsymbol I}\right)+\mathcal{H}\left(\tau \right){\times }_{3}\left({\boldsymbol W}-{\boldsymbol I}\right)+\mathcal{H}\left(0\right) $$ \begin{aligned}&\mathcal{H}\left(\tau \right){\times }_{1}\widehat{{\boldsymbol A}}+\mathcal{H}\left(\tau \right){\times }_{2}{\boldsymbol U}{\times }_{3}{\boldsymbol W} -\\&\quad \mathcal{H}\left(\tau \right){\times }_{2}{\boldsymbol I}{\times }_{3}{\boldsymbol W}-\mathcal{H}\left(\tau \right){\times }_{2}{\boldsymbol U}\times {\boldsymbol I}\end{aligned} $
MTGODE[47]$ \dfrac{\mathrm{d}{{\boldsymbol H}}_{{T}}^{t}}{\mathrm{d}t}=\mathcal{P}\left(\mathcal{A}\left({{O}{D}{E}{S}{o}{l}{v}{e}}^{2}\left({T}{C}{N}\left({{\boldsymbol H}}_{{T}}^{t},t,{\boldsymbol \theta }\right), \dfrac{\mathrm{d}{{\boldsymbol H}}_{\mathrm{G}}^{\tau }}{\mathrm{d}\tau },0,… ,T\right),{\boldsymbol \varPhi }\right),R\right) $$ {T}{C}{N}\left({{\boldsymbol H}}_{{T}}^{t},t,{\boldsymbol \theta }\right) $
ASTGCN[59]$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }={\boldsymbol H}\left(\tau \right){\times }_{1}\left(\dfrac{\alpha }{2}{\widehat{\boldsymbol A}}-{{\boldsymbol I}}_{n}\right)+{\boldsymbol H}\left(\tau \right){\times }_{2}\left({\boldsymbol U}-{{\boldsymbol I}}_{n}\right)+{\boldsymbol H}\left(\tau \right){\times }_{3}\left({\boldsymbol W}-{{\boldsymbol I}}_{n}\right)+{\boldsymbol H}\left(0\right) $$ \begin{aligned}&{\boldsymbol H}\left(\tau \right){\times }_{1}\dfrac{\alpha }{2}{\widehat{\boldsymbol A}}+{\boldsymbol H}\left(\tau \right){\times }_{2}{\boldsymbol U}{\times }_{3}{\boldsymbol W} -\\&\quad {\boldsymbol H}\left(\tau \right){\times }_{2}{{\boldsymbol I}}_{n}{\times }_{3}{\boldsymbol W}-{\boldsymbol H}\left(\tau \right){\times }_{2}{\boldsymbol U}\times {{\boldsymbol I}}_{n}\end{aligned} $
CG-ODE[48]$ \dfrac{{\rm d}{{\boldsymbol H}}^{t}}{{\rm d}t}=\sigma \left({{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{\boldsymbol W}\right)-{{\boldsymbol H}}^{t}+{{\boldsymbol H}}^{0} $$ \sigma \left({{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{\boldsymbol W}\right) $
HOPE[61]$ \dfrac{{\mathrm{d}}^{2}{{\boldsymbol H}}^{t}}{\mathrm{d}{t}^{2}}+\gamma \dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}=\sigma \left({\left({{\widehat{\boldsymbol D}}}^{t}\right)}^{-1}{{{{\widehat{\boldsymbol A}}_{\mathrm{U}}}}^{\mathit{t}}{\boldsymbol H}}^{t}{{\boldsymbol W}}_{p}\right)-{{\boldsymbol H}}^{t} $$ \sigma \left({\left({{\widehat{\boldsymbol D}}}^{t}\right)}^{-1}{{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{{\boldsymbol W}}_{p}\right) $
SEGODE[27]$ \dfrac{{\rm d}{{\boldsymbol H}}^{t}}{{\rm d}t}={s}{o}{f}{t}{m}{a}{x}{\left(\dfrac{{{\boldsymbol Q}}^{t}{{\boldsymbol K}}^{t}}{\sqrt{{d}_{k}}}\right)^{T}}{{\boldsymbol V}}^{t} $$ {A}{T}{T}{E}{N}{T}{I}{O}{N}({{\boldsymbol H}}^{t},t,\boldsymbol{\varTheta }) $
GNG-ODE[60]$ \dfrac{\rm{d}\boldsymbol{h}_v^t}{\rm{d}t}=(1-\boldsymbol{z}_v^t)\odot(\boldsymbol{g}_{\mathit{v}}^t-\boldsymbol{h}_v^t) $$ {G}{R}{U}({{\boldsymbol H}}^{t},t,\boldsymbol{\varTheta }) $
), ArticleFig(id=1329457296918295345, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表4, caption=

离散动态图神经ODE

, figureFileSmall=null, figureFileBig=null, tableContent=
算法ODE函数$ {\widehat{\boldsymbol H}}^{t} $ODE作用
NDCN[31]$ \dfrac{{\rm d}{\boldsymbol H'^{t}}}{{\rm d}t}=f\left({\boldsymbol H'},G,{{\boldsymbol W}}_{\mathrm{h}},t\right) $$ {f}_{\mathrm{E}}\left({{\boldsymbol H}}^{\mathit{t}},{{\boldsymbol W}}_{\mathrm{E}}\right) $推断器
STG-ODE[46]$ \dfrac{\mathrm{d}\mathcal{H}\left(\tau \right)}{\mathrm{d}\tau }=\mathcal{H}\left(\tau \right){\times }_{1}\left({\widehat{\boldsymbol A}}-{\boldsymbol I}\right)+\mathcal{H}\left(\tau \right){\times }_{2}\left({\boldsymbol U}-{\boldsymbol I}\right)+\mathcal{H}\left(\tau \right){\times }_{3}\left({\boldsymbol W}-{\boldsymbol I}\right)+\mathcal{H}\left(0\right) $$ \begin{aligned}&\mathcal{H}\left(\tau \right){\times }_{1}\widehat{{\boldsymbol A}}+\mathcal{H}\left(\tau \right){\times }_{2}{\boldsymbol U}{\times }_{3}{\boldsymbol W} -\\&\quad \mathcal{H}\left(\tau \right){\times }_{2}{\boldsymbol I}{\times }_{3}{\boldsymbol W}-\mathcal{H}\left(\tau \right){\times }_{2}{\boldsymbol U}\times {\boldsymbol I}\end{aligned} $
MTGODE[47]$ \dfrac{\mathrm{d}{{\boldsymbol H}}_{{T}}^{t}}{\mathrm{d}t}=\mathcal{P}\left(\mathcal{A}\left({{O}{D}{E}{S}{o}{l}{v}{e}}^{2}\left({T}{C}{N}\left({{\boldsymbol H}}_{{T}}^{t},t,{\boldsymbol \theta }\right), \dfrac{\mathrm{d}{{\boldsymbol H}}_{\mathrm{G}}^{\tau }}{\mathrm{d}\tau },0,… ,T\right),{\boldsymbol \varPhi }\right),R\right) $$ {T}{C}{N}\left({{\boldsymbol H}}_{{T}}^{t},t,{\boldsymbol \theta }\right) $
ASTGCN[59]$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }={\boldsymbol H}\left(\tau \right){\times }_{1}\left(\dfrac{\alpha }{2}{\widehat{\boldsymbol A}}-{{\boldsymbol I}}_{n}\right)+{\boldsymbol H}\left(\tau \right){\times }_{2}\left({\boldsymbol U}-{{\boldsymbol I}}_{n}\right)+{\boldsymbol H}\left(\tau \right){\times }_{3}\left({\boldsymbol W}-{{\boldsymbol I}}_{n}\right)+{\boldsymbol H}\left(0\right) $$ \begin{aligned}&{\boldsymbol H}\left(\tau \right){\times }_{1}\dfrac{\alpha }{2}{\widehat{\boldsymbol A}}+{\boldsymbol H}\left(\tau \right){\times }_{2}{\boldsymbol U}{\times }_{3}{\boldsymbol W} -\\&\quad {\boldsymbol H}\left(\tau \right){\times }_{2}{{\boldsymbol I}}_{n}{\times }_{3}{\boldsymbol W}-{\boldsymbol H}\left(\tau \right){\times }_{2}{\boldsymbol U}\times {{\boldsymbol I}}_{n}\end{aligned} $
CG-ODE[48]$ \dfrac{{\rm d}{{\boldsymbol H}}^{t}}{{\rm d}t}=\sigma \left({{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{\boldsymbol W}\right)-{{\boldsymbol H}}^{t}+{{\boldsymbol H}}^{0} $$ \sigma \left({{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{\boldsymbol W}\right) $
HOPE[61]$ \dfrac{{\mathrm{d}}^{2}{{\boldsymbol H}}^{t}}{\mathrm{d}{t}^{2}}+\gamma \dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}=\sigma \left({\left({{\widehat{\boldsymbol D}}}^{t}\right)}^{-1}{{{{\widehat{\boldsymbol A}}_{\mathrm{U}}}}^{\mathit{t}}{\boldsymbol H}}^{t}{{\boldsymbol W}}_{p}\right)-{{\boldsymbol H}}^{t} $$ \sigma \left({\left({{\widehat{\boldsymbol D}}}^{t}\right)}^{-1}{{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{{\boldsymbol W}}_{p}\right) $
SEGODE[27]$ \dfrac{{\rm d}{{\boldsymbol H}}^{t}}{{\rm d}t}={s}{o}{f}{t}{m}{a}{x}{\left(\dfrac{{{\boldsymbol Q}}^{t}{{\boldsymbol K}}^{t}}{\sqrt{{d}_{k}}}\right)^{T}}{{\boldsymbol V}}^{t} $$ {A}{T}{T}{E}{N}{T}{I}{O}{N}({{\boldsymbol H}}^{t},t,\boldsymbol{\varTheta }) $
GNG-ODE[60]$ \dfrac{\rm{d}\boldsymbol{h}_v^t}{\rm{d}t}=(1-\boldsymbol{z}_v^t)\odot(\boldsymbol{g}_{\mathit{v}}^t-\boldsymbol{h}_v^t) $$ {G}{R}{U}({{\boldsymbol H}}^{t},t,\boldsymbol{\varTheta }) $
), ArticleFig(id=1329457296989598514, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 5, caption=

Continuous Dynamic Graph Neural ODE

, figureFileSmall=null, figureFileBig=null, tableContent=
算法ODE函数消息聚合ODE作用
ConTIG[32]$ \dfrac{{\rm d}{{\boldsymbol H}}^{{t}_{p}}\left(\tau \right)}{\mathrm{d}\tau }={z}_{l}\odot {{\boldsymbol H}}^{{t}_{p}}\left(0\right)+{z}_{n}\odot {{\boldsymbol A}}_{p}{{\boldsymbol H}}^{{t}_{p}}\left(\tau \right)-{z}_{i}\odot {{\boldsymbol H}}^{{t}_{p}}\left(\tau \right) $$ {F}{C}\left({{\boldsymbol h }}_{v}^{{t}_{p-1}}\| {{\boldsymbol h }}_{u}^{{t}_{p-1}}\| {f}^{vu}\| {F}_{T}\left(\Delta t\right)\right) $编码器
NeurTWs[62]$ \dfrac{\mathrm{d}{{\boldsymbol H}}_{i}}{\mathrm{d}t}={{\boldsymbol H}}_{i-1}+\displaystyle\int _{{t}_{i-1}}^{{t}_{i}}f\left({{\boldsymbol H}}_{t},{\boldsymbol \theta} \right){\rm d}t $$ \begin{aligned} &{A'}\left({w}_{i}\right)={M}{L}{P}\left(A\left({w}_{i}\right),\psi \right)\\& {{\boldsymbol H}}_{i}={R}{N}{N}\left(\dfrac{\mathrm{d}{{\boldsymbol H}}_{i}}{\mathrm{d}t},{A'}\left({w}_{i}\right),\varphi \right)\end{aligned} $
GDERec[63]$ \begin{aligned} &\dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}=\mathrm{l}\mathrm{n}{\boldsymbol A}{{\boldsymbol A}}^{t}{{\boldsymbol H}}^{0}+{{\boldsymbol A}}^{t+1}{{\boldsymbol H}}^{0}\odot {{\boldsymbol H}}^{0}\\&\dfrac{{\mathrm{d}}^{2}{{\boldsymbol H}}^{t}}{\mathrm{d}{t}^{2}}={\mathrm{l}\mathrm{n}}^{2}{\boldsymbol A}{{\boldsymbol A}}^{t}{{\boldsymbol H}}^{0}+\mathrm{l}\mathrm{n}{\boldsymbol A}{{\boldsymbol A}}^{t+1}{{\boldsymbol H}}^{0}\odot {{\boldsymbol H}}^{0} =\mathrm{l}\mathrm{n}{\boldsymbol A}\dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}\end{aligned} $$ {{\boldsymbol A}}^{t+1}{{\boldsymbol H}}^{0}\odot {{\boldsymbol H}}^{0} $
GCDE-RNN[55]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}}}{\mathrm{d}t}={G}{C}{N}\left({R}{N}{N}\right({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}}),{\boldsymbol \varTheta }) $$ {R}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}}) $
GNODEe[64]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}}}{\mathrm{d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{O}\mathrm{D}\mathrm{E}-\mathrm{G}\mathrm{R}\mathrm{N}\mathrm{N}}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{\boldsymbol{G}}^{{\mathit{t}}^{\mathit{i}}},{\boldsymbol \varTheta }) $$ {G}{R}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{\boldsymbol{G}}^{{\mathit{t}}^{\mathit{i}}}) $
GNODEi[64]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}={G}{C}{N}({{\boldsymbol H}}^{t},{\tilde{D}}_{g}({{\boldsymbol H}}^{t})) $$ {\tilde{D}}_{g}({{\boldsymbol H}}^{t}) $推断器
LG-ODE[56]$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t}}{{\rm d}t}=f\left({{\boldsymbol h }}_{{v}_{1}}^{t},{{\boldsymbol h }}_{{v}_{2}}^{t},…, {{\boldsymbol h }}_{{v}_{N}}^{t}\right)={f}_{O}\left(\displaystyle\sum _{j\in {\mathcal{N}}_{i}}{f}_{R}\left(\left[{{\boldsymbol h }}_{{v}_{i}}^{t},{{\boldsymbol h }}_{{v}_{j}}^{t}\right]\right)\right) $$ \begin{aligned} &{{\boldsymbol z}}_{i}^{t}={f}_{\mathrm{E}}\left({v}_{i},{v}_{j}|\mathrm{i}\mathrm{f}\ {v}_{j}\in \mathcal{N}\left({v}_{i}\right)\right) \\&{{\boldsymbol u}}_{i}={f}_{\text{аggre}}\left({{\boldsymbol z}}_{i}^{{t}_{1}},{{\boldsymbol z}}_{i}^{{t}_{2}},…,{{\boldsymbol z}}_{i}^{{t}_{{T}_{i}}}\right)\end{aligned}$
TANGO[52]$ \begin{aligned} &\dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}}{{\rm d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({{\boldsymbol h }}_{{v}_{1}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}},{{\boldsymbol h }}_{{v}_{2}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}},…, {{\boldsymbol h }}_{{v}_{N}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}\right) \\&\dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}}}{{\rm d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({{\boldsymbol h }}_{{v}_{1}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}},{{\boldsymbol h }}_{{v}_{2}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}},…, {{\boldsymbol h }}_{{v}_{N}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}}\right)\end{aligned} $$ {{\boldsymbol h }}_{{v}_{0}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}={f}_{\mathrm{E}\mathrm{N}\mathrm{C}}({\boldsymbol X}\left({t}_{1}:K\right),\mathcal{G}) $
GG-ODE[57]$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,e}}{{\rm d}t}=\displaystyle\sum _{j\in {\mathcal{N}}_{i}}{f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({\hat{{\boldsymbol h }}}_{{v}_{i}}^{t,e},{\hat{{\boldsymbol h }}}_{{v}_{j}}^{t,e}\right)+{f}_{\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{f}}\left({\hat{{\boldsymbol h }}}_{{v}_{i}}^{t,e}\right) $$ {{\widehat{\boldsymbol h }}}_{{v}_{i}}^{t,e}={f}_{\mathrm{e}\mathrm{n}\mathrm{v}}\left({{\boldsymbol h }}_{{v}_{i}}^{t,e}\|{{\boldsymbol u}}^{e}\right) $
CF-GODE[50]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}}}{\mathrm{d}t}={G}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}};{\boldsymbol \varTheta }) $$ {G}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}};{\boldsymbol \varTheta }) $
), ArticleFig(id=1329457297073484595, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表5, caption=

连续动态图神经ODE

, figureFileSmall=null, figureFileBig=null, tableContent=
算法ODE函数消息聚合ODE作用
ConTIG[32]$ \dfrac{{\rm d}{{\boldsymbol H}}^{{t}_{p}}\left(\tau \right)}{\mathrm{d}\tau }={z}_{l}\odot {{\boldsymbol H}}^{{t}_{p}}\left(0\right)+{z}_{n}\odot {{\boldsymbol A}}_{p}{{\boldsymbol H}}^{{t}_{p}}\left(\tau \right)-{z}_{i}\odot {{\boldsymbol H}}^{{t}_{p}}\left(\tau \right) $$ {F}{C}\left({{\boldsymbol h }}_{v}^{{t}_{p-1}}\| {{\boldsymbol h }}_{u}^{{t}_{p-1}}\| {f}^{vu}\| {F}_{T}\left(\Delta t\right)\right) $编码器
NeurTWs[62]$ \dfrac{\mathrm{d}{{\boldsymbol H}}_{i}}{\mathrm{d}t}={{\boldsymbol H}}_{i-1}+\displaystyle\int _{{t}_{i-1}}^{{t}_{i}}f\left({{\boldsymbol H}}_{t},{\boldsymbol \theta} \right){\rm d}t $$ \begin{aligned} &{A'}\left({w}_{i}\right)={M}{L}{P}\left(A\left({w}_{i}\right),\psi \right)\\& {{\boldsymbol H}}_{i}={R}{N}{N}\left(\dfrac{\mathrm{d}{{\boldsymbol H}}_{i}}{\mathrm{d}t},{A'}\left({w}_{i}\right),\varphi \right)\end{aligned} $
GDERec[63]$ \begin{aligned} &\dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}=\mathrm{l}\mathrm{n}{\boldsymbol A}{{\boldsymbol A}}^{t}{{\boldsymbol H}}^{0}+{{\boldsymbol A}}^{t+1}{{\boldsymbol H}}^{0}\odot {{\boldsymbol H}}^{0}\\&\dfrac{{\mathrm{d}}^{2}{{\boldsymbol H}}^{t}}{\mathrm{d}{t}^{2}}={\mathrm{l}\mathrm{n}}^{2}{\boldsymbol A}{{\boldsymbol A}}^{t}{{\boldsymbol H}}^{0}+\mathrm{l}\mathrm{n}{\boldsymbol A}{{\boldsymbol A}}^{t+1}{{\boldsymbol H}}^{0}\odot {{\boldsymbol H}}^{0} =\mathrm{l}\mathrm{n}{\boldsymbol A}\dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}\end{aligned} $$ {{\boldsymbol A}}^{t+1}{{\boldsymbol H}}^{0}\odot {{\boldsymbol H}}^{0} $
GCDE-RNN[55]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}}}{\mathrm{d}t}={G}{C}{N}\left({R}{N}{N}\right({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}}),{\boldsymbol \varTheta }) $$ {R}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}}) $
GNODEe[64]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}}}{\mathrm{d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{O}\mathrm{D}\mathrm{E}-\mathrm{G}\mathrm{R}\mathrm{N}\mathrm{N}}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{\boldsymbol{G}}^{{\mathit{t}}^{\mathit{i}}},{\boldsymbol \varTheta }) $$ {G}{R}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{\boldsymbol{G}}^{{\mathit{t}}^{\mathit{i}}}) $
GNODEi[64]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}={G}{C}{N}({{\boldsymbol H}}^{t},{\tilde{D}}_{g}({{\boldsymbol H}}^{t})) $$ {\tilde{D}}_{g}({{\boldsymbol H}}^{t}) $推断器
LG-ODE[56]$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t}}{{\rm d}t}=f\left({{\boldsymbol h }}_{{v}_{1}}^{t},{{\boldsymbol h }}_{{v}_{2}}^{t},…, {{\boldsymbol h }}_{{v}_{N}}^{t}\right)={f}_{O}\left(\displaystyle\sum _{j\in {\mathcal{N}}_{i}}{f}_{R}\left(\left[{{\boldsymbol h }}_{{v}_{i}}^{t},{{\boldsymbol h }}_{{v}_{j}}^{t}\right]\right)\right) $$ \begin{aligned} &{{\boldsymbol z}}_{i}^{t}={f}_{\mathrm{E}}\left({v}_{i},{v}_{j}|\mathrm{i}\mathrm{f}\ {v}_{j}\in \mathcal{N}\left({v}_{i}\right)\right) \\&{{\boldsymbol u}}_{i}={f}_{\text{аggre}}\left({{\boldsymbol z}}_{i}^{{t}_{1}},{{\boldsymbol z}}_{i}^{{t}_{2}},…,{{\boldsymbol z}}_{i}^{{t}_{{T}_{i}}}\right)\end{aligned}$
TANGO[52]$ \begin{aligned} &\dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}}{{\rm d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({{\boldsymbol h }}_{{v}_{1}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}},{{\boldsymbol h }}_{{v}_{2}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}},…, {{\boldsymbol h }}_{{v}_{N}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}\right) \\&\dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}}}{{\rm d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({{\boldsymbol h }}_{{v}_{1}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}},{{\boldsymbol h }}_{{v}_{2}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}},…, {{\boldsymbol h }}_{{v}_{N}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}}\right)\end{aligned} $$ {{\boldsymbol h }}_{{v}_{0}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}={f}_{\mathrm{E}\mathrm{N}\mathrm{C}}({\boldsymbol X}\left({t}_{1}:K\right),\mathcal{G}) $
GG-ODE[57]$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,e}}{{\rm d}t}=\displaystyle\sum _{j\in {\mathcal{N}}_{i}}{f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({\hat{{\boldsymbol h }}}_{{v}_{i}}^{t,e},{\hat{{\boldsymbol h }}}_{{v}_{j}}^{t,e}\right)+{f}_{\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{f}}\left({\hat{{\boldsymbol h }}}_{{v}_{i}}^{t,e}\right) $$ {{\widehat{\boldsymbol h }}}_{{v}_{i}}^{t,e}={f}_{\mathrm{e}\mathrm{n}\mathrm{v}}\left({{\boldsymbol h }}_{{v}_{i}}^{t,e}\|{{\boldsymbol u}}^{e}\right) $
CF-GODE[50]$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}}}{\mathrm{d}t}={G}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}};{\boldsymbol \varTheta }) $$ {G}{N}{N}({{\boldsymbol H}}^{{\mathit{t}}^{\mathit{i}}},{{\boldsymbol X}}^{{\mathit{t}}^{\mathit{i}}};{\boldsymbol \varTheta }) $
), ArticleFig(id=1329457297199313717, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 6, caption=

Statistics of Node Classification Datasets

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集节点数边数特征数分类数
Cora2 7085 4291 4337
Citeseer3 3274 7323 7036
Pubmed19 71744 3385003
NELL65 755266 1445 414210
), ArticleFig(id=1329457297287394101, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表6, caption=

节点分类数据集统计

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集节点数边数特征数分类数
Cora2 7085 4291 4337
Citeseer3 3274 7323 7036
Pubmed19 71744 3385003
NELL65 755266 1445 414210
), ArticleFig(id=1329457297346114358, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 7, caption=

Statistics of Link Prediction Datasets

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集节点数边数时间粒度特征数
Enron14362 617s0/0
CollegeMsg1 89959 835s0/0
Mooc7 144411 749s0/4
Wikipedia9 227157 474s172/172
Reddit10 984672 447s172/172
Taobao64 70377 436s0/4
), ArticleFig(id=1329457297417417528, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表7, caption=

链接预测数据集统计

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集节点数边数时间粒度特征数
Enron14362 617s0/0
CollegeMsg1 89959 835s0/0
Mooc7 144411 749s0/4
Wikipedia9 227157 474s172/172
Reddit10 984672 447s172/172
Taobao64 70377 436s0/4
), ArticleFig(id=1329457297488720697, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=EN, label=Table 8, caption=

Sequence Prediction Datasets

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集节点数边数时间跨度/d
PeMSD7(M)2281 13212 672
PeMSD7(L)1 02610 15012 672
PeMS0335854726 208
PeMS0430734016 992
PeMS0788386628 224
PeMS0817029517 856
), ArticleFig(id=1329457297551635258, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, language=CN, label=表8, caption=

序列预测数据集

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集节点数边数时间跨度/d
PeMSD7(M)2281 13212 672
PeMSD7(L)1 02610 15012 672
PeMS0335854726 208
PeMS0430734016 992
PeMS0788386628 224
PeMS0817029517 856
)], attaches=null, journal=Journal(id=1324759606947315908, delFlag=0, nameCn=计算机研究与发展, nameEn=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, issn=1000-1239, eissn=null, cn=11-1777/TP, coden=null, periodic=0, language=CN, oaType=null, ccby=null, superviseOffice=null, ownerOffice=null, pubOffice=null, editorOffice=null, officeType=null, aims=null, clcCode=null, officeProv=null, officeCity=null, officeAddr=null, officeZip=null, officeEmail=null, officePhone=null, editDirector=null, officeDirector=null, officeDirectorPhone=null, officeStaffNum=null, officeEmpNum=null, coverPicUrl=k2l8zgqwstuxxevFX2qxVA==, journalPrice=null, startedYear=null, abbrevIsoEn=Journal of Computer Research and Development, journalRemark=null, publicationField=null, createdTime=1793776903141, updatedTime=1793850698612, createdBy=18614031015, updatedBy=13701087609, firstLetterCn=J, firstLetterEn=J, subjectCode=Electronic Information, subjectName=null, subjectCodeEn=Electronic Information, subjectNameEn=null, picCn=k2l8zgqwstuxxevFX2qxVA==, picEn=r6k0Ry+RD+c/VXj5QarmDw==, jcr=null, cjcr=null, exts=[JournalExt(id=1325069127710286688, language=CN, name=计算机研究与发展, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698640, updatedTime=1793850698640, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""}), JournalExt(id=1325069127777395553, language=EN, name=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698656, updatedTime=1793850698656, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""})], databaseList=null, tenantJournalId=1324759914146529478, websiteList=[Website(id=1325070300139892771, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN, language=CN, createTime=1793850978169, createBy=18614031015, updateTime=1793851040259, updateBy=18614031015, name=计算机研究与发展-中文, tplId=1146099689490845704, title=计算机研究与发展, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070893751349360, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=articleTextType, value=kx, createTime=1793851119697, updateTime=1793851119697, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893730377837, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=banner, value=null, createTime=1793851119692, updateTime=1793851119692, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893772320883, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=grayFlag, value=0, createTime=1793851119702, updateTime=1793851119702, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893721989228, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851119690, updateTime=1793851119690, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893784903797, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=minRunFlag, value=0, createTime=1793851119705, updateTime=1793851119705, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893742960751, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic, createTime=1793851119695, updateTime=1793851119695, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893780709492, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=silenceFlag, value=0, createTime=1793851119704, updateTime=1793851119704, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893738766446, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_cn_619/, createTime=1793851119694, updateTime=1793851119694, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893755543665, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeColor, value=null, createTime=1793851119698, updateTime=1793851119698, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893763932274, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeStyle, value=null, createTime=1793851119700, updateTime=1793851119700, creator=18614031015, updator=18614031015)]), Website(id=1325070300580294698, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN, language=EN, createTime=1793850978274, createBy=18614031015, updateTime=1793851036638, updateBy=18614031015, name=计算机研究与发展-英文, tplId=1146101810881728533, title=Journal of Computer Research and Development, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070925288321148, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=articleTextType, value=kx, createTime=1793851127216, updateTime=1793851127216, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925263155321, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=banner, value=null, createTime=1793851127210, updateTime=1793851127210, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925309292671, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=grayFlag, value=0, createTime=1793851127221, updateTime=1793851127221, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925254766712, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851127208, updateTime=1793851127208, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925326069889, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=minRunFlag, value=0, createTime=1793851127225, updateTime=1793851127225, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925279932539, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic, createTime=1793851127214, updateTime=1793851127214, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925317681280, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=silenceFlag, value=0, createTime=1793851127223, updateTime=1793851127223, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925271543930, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_en_623/, createTime=1793851127212, updateTime=1793851127212, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925292515453, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeColor, value=null, createTime=1793851127217, updateTime=1793851127217, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925300904062, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeStyle, value=null, createTime=1793851127219, updateTime=1793851127219, creator=18614031015, updator=18614031015)])], journalTitle=计算机研究与发展, weixinUrl=null, journalUrl=https://crad.ict.ac.cn/, iacademicId=null, status=1, seqNo=null, journalTitleEn=Journal of Computer Research and Development, journalPhotoCn=k2l8zgqwstuxxevFX2qxVA==, journalPhotoEn=r6k0Ry+RD+c/VXj5QarmDw==, journalFirstLetter=J, journalRecommend=null, journalNew=null, journalCollection=null, jcrJf=null, cjcrJf=null, jcrJfStr=null, cjcrJfStr=null, submissionFirstDecision=null, sciSubjectClassification=null, casSubjectClassification=null, citeScore=null, totalCitationFrequency=null, icpCode=null, psCode=null, advertisingLicenseCode=null, copyrightInformation=null, country=null, option=, provinceCode=null, provinceName=null, collectFlag=false, interPubPlatform=, interPubPlatformUrl=null), detailUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202440192, detailUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/10.7544/issn1000-1239.202440192, pdfUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/PDF/10.7544/issn1000-1239.202440192, pdfUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/PDF/10.7544/issn1000-1239.202440192, aliStartDate=0, aliEndDate=0, collectionFlag=false, citedCount=null, citedUrl=null, previewStatus=0, delFlag=0, hasFullText=1, orderTime=1722441600000, fullTextJson=null, articleText=null, reference=null)
收藏切换
图神经常微分方程综述
焦鹏飞 1 , 陈舒欣 2 , 郭翾 3 , 何东晓 3, * , 刘栋 4
计算机研究与发展 | 图机器学习专题 2024,61(8): 2045-2066
收起
收藏切换
计算机研究与发展 |图机器学习专题 2024 , 61 (8) : 2045 -2066
图神经常微分方程综述
全屏
[Author(id=1329457288680682226, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=pjiao@hdu.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457288760374004, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288680682226, language=EN, stringName=Pengfei Jiao, firstName=Pengfei, middleName=null, lastName=Jiao, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Cyberspace, Hangzhou Dianzi University, Hangzhou 310018, bio={"content":"

born in 1990. PhD, professor. His main research interest includes complex network analysis and its applications

"}, bioImg=null, bioContent=

born in 1990. PhD, professor. His main research interest includes complex network analysis and its applications

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457288840065781, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288680682226, language=CN, stringName=焦鹏飞, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1杭州电子科技大学网络空间安全学院 杭州 310018, bio={"img":"xkFJCUTEussLht1CvuM/7w==","content":"

1990年生. 博士,教授. 主要研究方向为复杂网络分析及其应用

"}, bioImg=xkFJCUTEussLht1CvuM/7w==, bioContent=

1990年生. 博士,教授. 主要研究方向为复杂网络分析及其应用

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288340943588, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=1, ext=[AuthorCompanyExt(id=1329457288349332197, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288340943588, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Cyberspace, Hangzhou Dianzi University, Hangzhou 310018), AuthorCompanyExt(id=1329457288357720806, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288340943588, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1杭州电子科技大学网络空间安全学院 杭州 310018)])]), Author(id=1329457288919757559, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457292392641274, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288919757559, language=EN, stringName=Shuxin Chen, firstName=Shuxin, middleName=null, lastName=Chen, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2ZhuoYue Honors College, Hangzhou Dianzi University, Hangzhou 310018, bio={"content":"

born in 2004. Undergraduate. Her main research interest includes graph neural network

"}, bioImg=null, bioContent=

born in 2004. Undergraduate. Her main research interest includes graph neural network

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457292489110268, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457288919757559, language=CN, stringName=陈舒欣, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2杭州电子科技大学卓越学院 杭州 310018, bio={"img":"vecXV3jDToAN4du0WSeQOQ==","content":"

2004年生. 本科生. 主要研究方向为图神经网络

"}, bioImg=vecXV3jDToAN4du0WSeQOQ==, bioContent=

2004年生. 本科生. 主要研究方向为图神经网络

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288408052455, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=2, ext=[AuthorCompanyExt(id=1329457288412246760, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288408052455, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2ZhuoYue Honors College, Hangzhou Dianzi University, Hangzhou 310018), AuthorCompanyExt(id=1329457288420635369, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288408052455, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2杭州电子科技大学卓越学院 杭州 310018)])]), Author(id=1329457292560413438, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457292640105216, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292560413438, language=EN, stringName=Xuan Guo, firstName=Xuan, middleName=null, lastName=Guo, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3College of Intelligence and Computing, Tianjin University, Tianjin 300350, bio={"content":"

born in 1996. PhD candidate. His main research interest includes graph machine learning

"}, bioImg=null, bioContent=

born in 1996. PhD candidate. His main research interest includes graph machine learning

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457292723991297, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292560413438, language=CN, stringName=郭翾, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, address=3天津大学智能与计算学部 天津 300350, bio={"img":"reHGELmAdLzAZJBAkdfn+Q==","content":"

1996年生. 博士研究生. 主要研究方向为图机器学习

"}, bioImg=reHGELmAdLzAZJBAkdfn+Q==, bioContent=

1996年生. 博士研究生. 主要研究方向为图机器学习

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288491938538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=3, ext=[AuthorCompanyExt(id=1329457288500327147, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Intelligence and Computing, Tianjin University, Tianjin 300350), AuthorCompanyExt(id=1329457288512910060, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3天津大学智能与计算学部 天津 300350)])]), Author(id=1329457292786905860, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=hedongxiao@tju.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1329457292858209030, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292786905860, language=EN, stringName=Dongxiao He, firstName=Dongxiao, middleName=null, lastName=He, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, *, address=3College of Intelligence and Computing, Tianjin University, Tianjin 300350, bio={"content":"

born in 1984. PhD, professor. Her main research interests include analysis of complex networks, graph data mining, and graph neural network

"}, bioImg=null, bioContent=

born in 1984. PhD, professor. Her main research interests include analysis of complex networks, graph data mining, and graph neural network

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457292954678023, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457292786905860, language=CN, stringName=何东晓, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=3, *, address=3天津大学智能与计算学部 天津 300350, bio={"img":"s+rhBDgLapnAhFkIkLmH3w==","content":"

1984年生. 博士,教授. 主要研究方向为复杂网络分析、图数据挖掘、图神经网络

"}, bioImg=s+rhBDgLapnAhFkIkLmH3w==, bioContent=

1984年生. 博士,教授. 主要研究方向为复杂网络分析、图数据挖掘、图神经网络

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288491938538, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=3, ext=[AuthorCompanyExt(id=1329457288500327147, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3College of Intelligence and Computing, Tianjin University, Tianjin 300350), AuthorCompanyExt(id=1329457288512910060, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288491938538, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=3天津大学智能与计算学部 天津 300350)])]), Author(id=1329457293034369801, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, orderNo=4, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329457293135033100, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457293034369801, language=EN, stringName=Dong Liu, firstName=Dong, middleName=null, lastName=Liu, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=4, address=4College of Computer and Information Engineering, Henan Normal University, Xinxiang, Henan 453007, bio={"content":"

born in 1976. PhD, professor. His main research interests include educational data mining and complex network analysis

"}, bioImg=null, bioContent=

born in 1976. PhD, professor. His main research interests include educational data mining and complex network analysis

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329457293197947662, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, authorId=1329457293034369801, language=CN, stringName=刘栋, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=4, address=4河南师范大学计算机与信息工程学院 河南新乡 453007, bio={"img":"c+9oh/J1LNC/PRFDEP3NiA==","content":"

1976年生. 博士,教授. 主要研究方向为教育数据挖掘、复杂网络分析

"}, bioImg=c+9oh/J1LNC/PRFDEP3NiA==, bioContent=

1976年生. 博士,教授. 主要研究方向为教育数据挖掘、复杂网络分析

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329457288592601838, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, xref=4, ext=[AuthorCompanyExt(id=1329457288596796143, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288592601838, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=4College of Computer and Information Engineering, Henan Normal University, Xinxiang, Henan 453007), AuthorCompanyExt(id=1329457288605184752, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329457284658344645, companyId=1329457288592601838, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=4河南师范大学计算机与信息工程学院 河南新乡 453007)])])]
焦鹏飞1 , 陈舒欣2, 郭翾3, 何东晓3, * , 刘栋4
作者信息
  • 1杭州电子科技大学网络空间安全学院 杭州 310018
  • 2杭州电子科技大学卓越学院 杭州 310018
  • 3天津大学智能与计算学部 天津 300350
  • 4河南师范大学计算机与信息工程学院 河南新乡 453007
通讯作者:
何东晓()
作者简介:

1990年生. 博士,教授. 主要研究方向为复杂网络分析及其应用

2004年生. 本科生. 主要研究方向为图神经网络

1996年生. 博士研究生. 主要研究方向为图机器学习

1984年生. 博士,教授. 主要研究方向为复杂网络分析、图数据挖掘、图神经网络

1976年生. 博士,教授. 主要研究方向为教育数据挖掘、复杂网络分析

Survey on Graph Neural Ordinary Differential Equations
Pengfei Jiao1 , Shuxin Chen2, Xuan Guo3, Dongxiao He3, * , Dong Liu4
Affiliations
  • 1School of Cyberspace, Hangzhou Dianzi University, Hangzhou 310018
  • 2ZhuoYue Honors College, Hangzhou Dianzi University, Hangzhou 310018
  • 3College of Intelligence and Computing, Tianjin University, Tianjin 300350
  • 4College of Computer and Information Engineering, Henan Normal University, Xinxiang, Henan 453007
  • born in 1990. PhD, professor. His main research interest includes complex network analysis and its applications

    born in 2004. Undergraduate. Her main research interest includes graph neural network

    born in 1996. PhD candidate. His main research interest includes graph machine learning

    born in 1984. PhD, professor. Her main research interests include analysis of complex networks, graph data mining, and graph neural network

    born in 1976. PhD, professor. His main research interests include educational data mining and complex network analysis

出版时间: 2024-08-01 doi: 10.7544/issn1000-1239.202440192
文章导航
收藏切换

图神经网络(graph neural network,GNN)是处理图结构数据的强大工具,能够捕捉节点间的复杂关系和特征. 但GNN的离散架构导致其在表示图结构、建模图演化、适应不规则数据和计算开销等方面面临诸多挑战. 面对这些挑战,神经常微分方程(ordinary differential equation,ODE)由于能够模拟系统状态的连续变化,具备“连续深度”的编码和推断能力,被作为解决GNN面临挑战的全新方法而引入. 然而,神经ODE是为欧式结构数据设计的,无法直接捕捉图结构特性. 因此,提出了图神经ODE,一种将神经ODE与GNN结合的新架构,可以更好地适应图结构数据并充分利用其特性. 近年来,图神经ODE相关研究已经深入到图机器学习的各个方向中,引发了新的研究热潮. 在此背景下,适时地对图神经ODE研究前沿进行了系统性综述. 首先,回顾了GNN的关键优势和面临的诸多挑战,阐述了引入神经ODE并与GNN结合的理论基础和实践意义. 随后,详细介绍了图神经ODE的背景和基本概念,并提出了一种新颖的分类方法,在此基础上对当前的相关方法进行了全面描述. 然后,介绍了相关研究常用的验证方法,包括下游任务及数据集. 进一步,深入探讨了图神经ODE在多个实用领域上的应用. 最后,对图神经ODE面临的挑战和未来发展趋势进行了总结和展望.

图机器学习  /  图神经网络  /  神经常微分方程  /  分类体系  /  综述

Graph neural networks (GNNs) are powerful tools for handling graph-structured data, capable of capturing complex relationships and features among nodes. However, the discrete architecture of GNNs leads to numerous challenges in representing graph structures, modeling graph evolution, adapting to irregular data, and managing computational costs. In response to these challenges, neural ordinary differential equations (ODEs) have been introduced as a novel method to address the challenges faced by GNNs, as they can simulate the continuous evolution of system states, providing continuous deep encoding and inference capabilities. However, neural ODEs are designed for Euclidean structured data and cannot directly capture the characteristics of graphs. Therefore, researchers have proposed graph neural ODEs, a new type of architectures that combines neural ODEs with GNNs, which can better adapt to graph-structured data and fully utilize its characteristics. In recent years, research related to graph neural ODEs has delved into various directions of graph machine learning, sparking a new research trend. In this context, we systematically review the relevant research of graph neural ODEs in a timely manner. Firstly, we review the key advantages of GNNs and the challenges they face, and elucidate the theoretical basis and practical significance of introducing neural ODEs and combining them with GNNs. Subsequently, we elaborate on the background and basic concepts of graph neural ODEs, proposing a novel taxonomy, and comprehensively describe some important current methods on the taxonomy. Then, we introduce commonly used verification methods in related research, including downstream tasks and datasets. Furthermore, we delve into the applications of graph neural ODEs in multiple practical fields. Finally, we summarize and prospect the challenges and future development trends of graph neural ODEs.

graph machine learning  /  graph neural networks  /  neural ordinary differential equations  /  taxonomy  /  survey
焦鹏飞, 陈舒欣, 郭翾, 何东晓, 刘栋. 图神经常微分方程综述. 计算机研究与发展, 2024 , 61 (8) : 2045 -2066 . DOI: 10.7544/issn1000-1239.202440192
Pengfei Jiao, Shuxin Chen, Xuan Guo, Dongxiao He, Dong Liu. Survey on Graph Neural Ordinary Differential Equations[J]. Journal of Computer Research and Development, 2024 , 61 (8) : 2045 -2066 . DOI: 10.7544/issn1000-1239.202440192
图神经网络(graph neural network,GNN)已成为图数据处理和表示的一种新的范式[1-3]. GNN的核心思想是设计多个消息传递层,这些层通过聚合节点及其邻居的表示,迭代更新每个节点的表示[4-6]. 这种强调局部邻域关系的思想使得基于消息传递的GNN在处理图结构时表现出色. 在此基础上,为了应对现实系统中普遍存在的动态变化,一些融合循环神经网络(recurrent neural network,RNN)或注意力机制的GNN还能够捕获与理解图数据中的时序演化模式,其中包括VGRNN[7]等离散动态图方法和TGAT[8]等连续动态图方法. GNN支持端到端学习,可以直接从输入的图结构数据学习到用于诸如节点分类、链接预测、图分类等各种特定任务的输出. 由于其简单性和有效性,GNN被广泛应用于多种图领域,例如,社交网络分析[9-10]、推荐系统[11-13]、蛋白质结构预测[14-15]、交通流预测[16-18]等.
但是GNN的离散神经网络架构存在一些固有弊端:1)离散架构限制了GNN对图结构的表示能力,使得模型难以捕获图中的复杂关系:2)离散架构导致GNN对演化动力学建模的欠缺以及时间建模粒度不足限制了模型时序推断的能力;3)离散架构对数据的处理过程较为固定,缺乏适应数据变化的灵活性,难以处理不规则的观测;4)离散架构的专用设计和冗余参数使模型复杂化,导致更高的计算和内存开销. 在离散架构的约束下,为GNN设计的各类优化方案难以被综合运用.
首先,对图结构有限的表示能力是离散架构的GNN面临的关键问题,过平滑现象是该问题的典型代表[19]. 由若干消息传递层组成的GNN能够有效地扩张感受野并编码其中的信息. 然而,这种伴随着节点聚合更广泛邻域范围内的邻居信息,使得不同节点的表示向量趋向一致,难以支撑各类下游任务,此即过平滑现象. 现有研究常采用2种策略缓解该问题[20]:一种策略以DropEdge[21],PairNorm[22]等方法为代表,对训练过程进行正则化;另一种策略以JKNet[23]等方法为代表,将残差连接引入到GNN中. 尽管这2种策略能有效缓解过平滑问题,但不断扩张的感受野会带来巨大的计算开销,难以实际应用.
其次,离散架构的GNN对现实系统动力学的建模不足,使其缺乏支撑应用决策的可解释性. 无论自然的或是人造的,现实动态系统的演化往往符合特定的动力学机制[24-25],能否捕获这些演化规律,决定了模型推断系统未来状态的能力. 常见的GNN方法或是利用RNN等离散架构建模时序变化[7],或是忽略演化过程、直接综合历史信息做出预测[8],这些方法无法建模连续时间上的演化,更无法阐述现实系统中全局性的动力学机制[26].
再者,对图结构和演化的建模能力不足,进一步导致GNN难以应对现实中常见的不规则观测数据. 例如,当动态图数据中存在快照缺失时,离散化的架构往往因无法有效地处理缺失部分而失效[27]. 此外,当快照的长度不均匀,即数据采样间隔存在变化时,离散架构同样表现不足. 比如RNN应用于不规则时间序列的一个标准技巧是将时间划分为相等大小的间隔,并使用平均值来插补或聚合观测值,这样的处理会破坏信息,不利于推断系统未来状态.
最后,当处理大规模图时,GNN的计算和存储需求也会急剧增加,这限制了GNN在社交网络或知识图谱等大型网络上的应用效率. 面对这一挑战,一些方法采用采样或利用历史训练结果提升计算效率. 例如,GAS[28]在每次前向传播计算中仅采样一小部分的实体和其对应的单跳邻域实体进行计算,同时用历史训练中得到的表示作为补充信息,实现将整个计算图修剪为子树,从而适应大规模图. 但是通过这种方式学到的模型结果受采样样本质量的影响.
为了打破离散架构约束,研究者们开始探索新的模型架构. 神经常微分方程(ordinary differential equation,ODE)作为深度学习和微分方程交叉的产物,提供了一种描述系统状态随时间连续变化的创新方法. Chen等人[29]认为,残差网络、RNN解码器和标准化流等的迭代更新等实际上可以被视为连续变换的欧拉离散化. 因此,他们将连续变化近似分解为更小的变换步骤,提出了首个神经ODE模型. 因此,与传统深度固定的离散神经网络架构不同,神经ODE可以被看作是具有“连续深度”的网络.
神经ODE将离散架构连续化的做法为解决GNN面临的问题提供了新的方向. 首先,神经ODE的自适应调节连续深度能促进训练过程中梯度的有效传播,克服梯度消失的问题,并且作为残差网络的连续化,确保每层在继承前一层特征的基础上进行处理,有效避免了节点特征随网络深化而趋同的现象. 其次,将问题建模为微分方程可以让模型更容易集成物理法则或先验知识,有助于提高GNN的时序推断能力和可解释性. 另外,将神经网络与连续变换的微分方程相结合,通过连续的动态系统描述,实现在任意时间点求解系统动力学,提供了一种天然的方式来处理时间上的不连续或不规则采样. 最后,神经ODE的连续变换可以用相对较少的参数来实现与传统网络层相似的功能,提高了参数的使用效率,减少了过拟合的风险.
因此,探索如何将神经ODE的优势应用于图结构数据成为了研究的新方向. 一些研究开始将神经ODE与GNN结合,使得模型能够更灵活地适应复杂图结构和演化模式. Xhonneux等人[30]提出的CGNN的工作是在这方面的开创性尝试. CGNN将传统的离散GNN层转化为连续的ODE,首次实现了ODE与GNN的结合,极大缓解了GNN的过平滑问题,同时具有恒定且较小的内存成本. NDCN模型[31]则在连续深度上整合GNN层,在动力学视角下将连续深度解释为物理时间的连续性,增强了GNN模型的可解释性和灵活性. ConTIG[32]采用神经ODE学习节点间的交互模式和节点状态轨迹,从而捕获节点嵌入轨迹的连续动态演化.
尽管图神经ODE具有多样化的研究价值和应用价值,但目前仍处于发展的初步阶段,许多核心问题还未被突破,主要体现在3个方面:1)图神经ODE缺少全面且有针对性的论述. 尽管图神经ODE工作被不断提出,还未有工作对其进行完整梳理和明确论述. 2)图神经ODE缺少系统性的分类对比. 图数据包含静态图、离散动态图、连续动态图等不同形式,在这些图上适配的图神经ODE模型不尽相同,在不同的情境下ODE发挥的作用也各有侧重. 目前缺少针对不同图的特点以及ODE的特定功能的工作系统性综述这些模型的方法和原理. 同时,也几乎没有对于图神经ODE的任务与应用系统地梳理和分析. 3)缺少对各类图神经ODE的基线验证任务、数据集和代码等的整理,无法针对不同的任务提供便利的算法支持.
针对以上3个问题,本文对近些年的图神经ODE研究进行了综述. 具体地,首先,回顾了GNN的优势和面临的挑战,描述了引入神经ODE,并将其与GNN结合的动机与意义. 其次,讲述了图的相关概念,概述了各种GNN的方法,并对神经ODE的原理进行推导. 接着,提出一种新的方法分类体系作为理解和分析图神经ODE的理论依据,并在此基础上对当前一些代表性方法进行了系统性描述和比较. 此外,详细介绍了该领域中常用的验证任务和数据集,并集成了该领域相关工作的论文及代码. 最后,阐述了当前图神经ODE的应用,并对其面临的挑战和未来发展趋势进行了总结和展望,从而为今后的研究提供有价值的参考.
本文的主要贡献有3点:
1)首篇针对图神经ODE进行综述的论文,详细阐释了提出图神经ODE的动机,归纳了图神经ODE的原理、实现方法以及面临的挑战与未来展望.
2)提出了全新的方法分类体系,根据静态图、离散动态图和连续动态图等不同类型图的特点以及ODE发挥的不同作用,将图神经ODE的相关工作分为3类描述和比较,方便研究者系统性理解图神经ODE的原理和优势.
3)整理了图神经ODE相关研究的验证任务和数据集,并收集了不同方法的论文及代码,为研究者进行经验性研究提供便利.
本节简要介绍图神经ODE有关的基础概念及其符号表达. 本文常见符号总结于表1中.
图是描述实体交互关系的一种通用语言,根据其是否涉及时序变化通常分为静态图和动态图2种类型.
静态图不涉及时序变化,可以被表示为$ G= (\mathcal{V}, \mathcal{E},\mathcal{X},\ell ) $,其中$ \mathcal{V}=\left\{{v}_{1},… ,{v}_{N}\right\} $是节点集合,$ \left|\mathcal{V}\right|=N $为节点数;$ \mathcal{E}=\left\{{e}_{1},… ,{e}_{M}\right\} $是边集合,$ \left|\mathcal{E}\right|=M $为边数;$ \mathcal{X}= \left\{{{\boldsymbol x}}_{1},… ,{{\boldsymbol x}}_{N}\right\} $是节点特征集合,$ \ell =\left\{{{\boldsymbol l}}_{1},… ,{{\boldsymbol l}}_{M}\right\} $表示边特征集合. 通常使用邻接矩阵$ {\boldsymbol A}\in {\left\{\mathrm{0,1}\right\}}^{N\times N} $表示$ G $中节点间的连接关系:若$ {v}_{i} $和$ {v}_{j} $相邻,则$ {{\mathit A}}_{i,j}=1 $,否则$ {{\mathit A}}_{i,j}=0 $. 对角矩阵$ {\boldsymbol D} $是度矩阵,对角元素$ {{\boldsymbol D}}_{i,i}=\displaystyle\sum _{j}{{\mathit A}}_{i,j} $即为节点$ {v}_{i} $的度. 将$ \mathcal{X} $归纳为矩阵$ {\boldsymbol X}\in {\mathbb{R}}^{n\times d} $表示$ G $上的节点特征.
动态图涉及时序变化,根据表示形式可分为离散动态图和连续动态图. 离散动态图通常将时序关系按照等长的时间间隔划分为一组有序的静态图快照来表示,即:
$ {\mathcal{G}}_{\mathrm{D}}=\left\{{G}^{1},{G}^{2},… ,{G}^{T}\right\}. $
连续动态图则基于交互事件流进行表示,即:
$ {\mathcal{G}}_{\mathrm{C}}=\{\left({u}_{i},{v}_{i},{t}_{i}\right);i=\mathrm{1,2},… \}, $
其中$ {u}_{i} $和$ {v}_{i} $是参与第$ i $次交互的节点,$ {t}_{i} $是时间戳.
基于对图数据复杂关系深入理解和高效处理的需求,GNN应运而生,GNN的基本架构主要包括2个关键步骤:聚合和更新. 以静态图为例,聚合步骤旨在收集每个节点的邻居信息,以获取局部邻域的综合表示,可以定义为:
$ \boldsymbol{a}_v^{\left(l+1\right)}=AGGREGATE^{\left(l+1\right)}\left(\left\{\boldsymbol{h}_u^{\left(l\right)}:u\in\mathcal{N}\left(v\right)\right\}\right), $
其中$ {\boldsymbol{a}}_{v}^{\left(l+1\right)} $是聚合后得到的节点$ v $的邻居信息,$ {{AGGREGATE}}^{\left(l+1\right)} $是第$ l+1 $层的聚合函数,用于汇总邻居的信息. 例如,图注意力网络(graph attention network,GAT)[33]的聚合函数为:
$ \boldsymbol{a}_v^{\left(l+1\right)}=\displaystyle\sum_{u\in\mathcal{N}\left(v\right)}^{ }\alpha_{vu}\boldsymbol{W}\boldsymbol{h}_u^{\left(l\right)}, $
其中,$ {\boldsymbol W} $为可更新的参数矩阵,$ {\alpha }_{vu} $为学习得到的$ v $和$ u $之间的注意力系数,hu是节点v的邻居节点u在第l层的表示向量.
更新步骤使用聚合得到的邻居信息来更新每个节点的特征表示,可以定义为:
$ {{\boldsymbol h }}_{v}^{\left(l+1\right)}={{UPDATE}}^{\left(l+1\right)}\left({{\boldsymbol h }}_{v}^{\left(l\right)},{{\boldsymbol a}}_{v}^{\left(l+1\right)}\right), $
这里,$ {\boldsymbol h}_{v}^{\left(l+1\right)} $是节点$ v $在第$ l+1 $层的新表示向量,$ {{UPDATE}}^{\left(l+1\right)} $是第$ l+1 $层的更新函数,负责整合节点自身的特征$ {\boldsymbol h}_{v}^{\left(l\right)} $和聚合得到的邻居信息$ {\boldsymbol a}_{v}^{\left(l+1\right)} $,以产生新的节点表示. GAT模型的更新函数为:
$ {{\boldsymbol h }}_{v}^{\left(l+1\right)}=\sigma \left({\alpha }_{vu}{\boldsymbol W}{{\boldsymbol h }}_{u}^{\left(l\right)}+{{\boldsymbol a}}_{v}^{\left(l+1\right)}\right). $
在连续动态图[34-36]上,GNN的架构与在静态图上基本一致,只是在聚合和更新过程中引入了时间信息. 对于离散动态图[37-39],You等人[40]提出一个适配的扩展框架,即:
$ \hat{\boldsymbol{H}}_t^{\left(l+1\right)}=GNN^{\left(l+1\right)}\left(\boldsymbol{H}_t^{\left(l\right)}\right), $
$ \boldsymbol{H}_t^{\left(l+1\right)}=INFERENCE^{\left(l+1\right)}\left(\boldsymbol{H}_{t-1}^{\left(l+1\right)},\hat{\boldsymbol{H}}_t^{\left(l+1\right)}\right), $
$ {\hat{{\boldsymbol H}}}_{t}^{\left(l+1\right)} $表示快照$ t $更新前在第$ l+1 $层的表示,$ {{\boldsymbol H}}_{t}^{\left(l+1\right)} $表示快照$ t $更新后在第$ l+1 $层的表示,$ {INFERENCE}^{\left(l+1\right)} $是第$ l+1 $层的推断函数.
神经ODE的核心思想是通过神经网络参数化的ODE表征隐空间的连续动力学. 在神经ODE框架下,传统神经网络中的离散层被替换为一个对深度或时间参数连续建模的过程[29]. 假设随时间$ t $连续变化的状态为$ {\boldsymbol h }\left(t\right) $,这个变化过程由一个参数化的函数$ f $所控制,则神经ODE可以用方程来表示:
$ \begin{array}r\dfrac{\mathrm{d}\boldsymbol{h}\left(t\right)}{\mathrm{d}t}=f\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)\end{array}, $
其中$ {\boldsymbol h }\left(t\right) $是网络的状态. $ t $是“时间”或“深度”变量. $ f $是一个神经网络,描述了$ {\boldsymbol h } $如何随时间$ t $变化.
神经ODE的前向传播过程可以总结为3个步骤:首先,给定每一个输入样本状态的初始值$ {\boldsymbol h }\left(0\right) $. 接着,定义一个由神经网络$ {f}_{\boldsymbol \theta }\left({\boldsymbol h }\left(t\right),t\right) $参数化的函数,描述状态随时间如何变化. 最后,应用ODE的数值解法(如欧拉法、龙格库塔算法等)和使用求解器来解ODE的初值问题,即:
$ \boldsymbol{h}(T)=\boldsymbol{h}(0)+\displaystyle\int_0^Tf\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)\mathrm{d}t, $
$ {\boldsymbol h }\left(T\right)={O}{D}{E}{S}{o}{l}{v}{e}\left({\boldsymbol h }\left(0\right),f,\left[0,T\right],{\boldsymbol \theta }\right). $
在神经ODE中,利用伴随灵敏度方法[41]对反向传播进行计算. 该方法的核心在于把反向传播过程重新构建为一个新的ODE初值问题,通过ODE求解器直接计算所需的梯度. 这种方法允许模型不必保存过程中的所有中间状态,从而高效计算梯度. 由此,可以借助可微分的损失函数来计算损失,从而优化模型性能. 假设输出状态是$ {\boldsymbol h }\left({t}_{1}\right) $,则有:
$ \begin{split}\mathcal{L}\left(\boldsymbol{h}\left(t_1\right)\right)= & \mathcal{L}(ODESolve\left(\boldsymbol{h}\left(0\right),f,t_0,t_1,\boldsymbol{\theta}\right)= \\ &\mathcal{L}\left(\boldsymbol{h}\left(t_0\right)+\displaystyle\int_{t_0}^{t_1}f_{\boldsymbol{\theta}}\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)\mathrm{d}t\right),\end{split}$
其中,$ \mathcal{L}(\cdot ) $表示损失函数. 定义伴随状态$ {\boldsymbol s}\left(t\right) $来表示损失如何依赖于某时刻的隐藏状态:
$ {\boldsymbol s}\left(t\right)=\dfrac{\mathrm{d}\mathcal{L}}{\mathrm{d}{\boldsymbol h }\left(t\right)}. $
通过伴随灵敏度法可以推导出描述随时间变化的伴随状态的微分方程:
$ \dfrac{\mathrm{d}\boldsymbol{s}\left(t\right)}{\mathrm{d}t}=-\boldsymbol{s}\left(t\right)\dfrac{\partial f_{\boldsymbol{\theta}}\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)}{\partial\boldsymbol{h}\left(t\right)}. $
因此,
$ \boldsymbol{s}\left(t_0\right)=\boldsymbol{s}\left(t_1\right)+\displaystyle\int_{t_1}^{t_0}\dfrac{\mathrm{d}\boldsymbol{s}\left(t\right)}{\mathrm{d}t}\mathrm{d}t=\boldsymbol{s}\left(t_1\right)-\displaystyle\int_{t_0}^{t_1}\boldsymbol{s}\left(t\right)\dfrac{\partial f_{\boldsymbol{\theta}}\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)}{\partial\boldsymbol{h}\left(t\right)}\mathrm{d}t. $
可定义:
$ \boldsymbol{s}_{\boldsymbol{\theta}}\left(t\right)=\dfrac{\mathrm{d}\mathcal{L}}{\mathrm{d}\boldsymbol{\theta}\left(t\right)}\text{,}\; \; \boldsymbol{s}_t\left(t\right)=\dfrac{\mathrm{d}\mathcal{L}}{\mathrm{d}t}. $
因为$ {\boldsymbol \theta } $不随时间变化,假定:
$ \dfrac{\partial {\boldsymbol \theta }\left(t\right)}{\partial t}=0\text{,}\dfrac{\mathrm{d}t}{\mathrm{d}t}=1. $
相似地可以推导出以下的伴随状态方程:
$ \dfrac{\mathrm{d}\boldsymbol{s}_{\boldsymbol{\theta}}\left(t\right)}{\mathrm{d}t}=-\boldsymbol{s}\left(t\right)\dfrac{\partial f\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)}{\partial\boldsymbol{\theta}\left(t\right)}. $
从而,
$ \boldsymbol{a}_t\left(t_1\right)=\dfrac{\mathrm{d}\mathcal{L}}{\mathrm{d}\boldsymbol{h}\left(t_1\right)}\dfrac{\mathrm{d}\boldsymbol{h}\left(t_1\right)}{\mathrm{d}t_1}=\boldsymbol{a}\left(t_1\right)f\left(\boldsymbol{h}\left(t_1\right),t_1,\boldsymbol{\theta}\right), $
${{\boldsymbol a}}_{t}\left({t}_{0}\right) ={{\boldsymbol a}}_{t}\left({t}_{1}\right)-\displaystyle\int _{{t}_{1}}^{{t}_{0}}{\boldsymbol a}\left(t\right)\dfrac{\partial f\left({\boldsymbol h }\left(t\right),t,{\boldsymbol \theta }\right)}{\partial t}\mathrm{d}t. $
将上述问题向量化,可以整合这些伴随方程,这样就可以使用单一的ODE求解器同时求解所有需要参与模型训练的参数的梯度. 可以设置$ {\boldsymbol s}\left(t\right) $,$ {{\boldsymbol s}}_{{\boldsymbol \theta }}\left(t\right) $和$ {{\boldsymbol s}}_{t}\left(t\right) $为:
$ \boldsymbol{s}_{\mathrm{aug}}=\left[\boldsymbol{s}\left(t\right),\boldsymbol{s}_{\boldsymbol{\theta}}\left(t\right),\boldsymbol{s}_t\left(t\right)\right], $
并推导出方程:
$ \dfrac{\mathrm{d}{{\boldsymbol s}}_{{\mathrm{aug}}}}{\mathrm{d}t}=-\left[{\boldsymbol s}\dfrac{\partial f}{\partial {\boldsymbol h }},{\boldsymbol s}\dfrac{\partial f}{\partial {\boldsymbol \theta }},{\boldsymbol s}\dfrac{\partial f}{\partial t}\right]. $
从而得到:
$ \boldsymbol{s}_{\boldsymbol{\theta}}\left(t\right)=-\boldsymbol{s}\left(t\right)\dfrac{\partial f\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)}{\partial\boldsymbol{\theta}}\text{,} $
进而:
$ \dfrac{\mathrm{d}\mathcal{L}}{\mathrm{d}\boldsymbol{\theta}}=\displaystyle\int_{t_0}^{t_1}\boldsymbol{s}\left(t\right)\dfrac{\partial f{_{\boldsymbol \theta}}\left(\boldsymbol{h}\left(t\right),t,\boldsymbol{\theta}\right)}{\partial\boldsymbol{\theta}}\mathrm{d}t. $
Rubanova等人[42]将RNN中的状态转换推广到由神经网络指定的连续时间动力学,提出了ODE-RNN模型. 其定义观测之间的状态变换为一个ODE问题:
$ \begin{array}{rr}\dfrac{{\rm d}{{\boldsymbol h }}_{i}}{{\rm d}t} ={ODESolve}\left({f}_{\boldsymbol \theta },{{\boldsymbol h }}_{i-1},\left({t}_{i-1},{t}_{i}\right)\right)\end{array},$
其中,$ {{\boldsymbol h }}_{i} $表示第$ i $个观测的表示. 在每次观测时,使用标准的RNN更新隐藏状态,即:
$ {{\boldsymbol h }}_{i}={RNN}\left(\dfrac{{\rm d}{{\boldsymbol h }}_{i}}{{\rm d}t},{{\boldsymbol X}}_{i}\right). $
此外,神经ODE在学习表示时保持了输入空间的拓扑结构,这导致它们无法表示某些特定类别的函数,只能学习与输入空间同胚的特征. 为解决这一问题,ANODE[43]通过扩展解决ODE的空间,避免轨迹相交,提出增广的ODE模型,即:
$ \dfrac{\mathrm{d}}{\mathrm{d}t}\left[ \begin{array}{c}\boldsymbol{h}\left(t\right) \\ \boldsymbol{a}\left(t\right)\end{array} \right]=f\left(\left[ \begin{array}{c}\boldsymbol{h}\left(t\right) \\ \boldsymbol{a}\left(t\right)\end{array} \right],t\right)\text{,}\; \; \left[ \begin{array}{c}\boldsymbol{h}\left(0\right) \\ \boldsymbol{a}\left(0\right)\end{array} \right]=\left[\begin{array}{c}\boldsymbol{x} \\ 0\end{array}\right], $
学习到更平滑的表示,也降低了计算内存.
ODE求解器从系统初始时间的已知状态开始,将时间分割成小的时间步长$ \mathrm{\Delta }t $,并在这些时间点上逐步近似求解ODE. 对于每个时间步长,求解器使用当前的状态和导数信息来估计下一个时间点的状态. 例如,求解器采用显式Euler方法[44]直接使用当前导数$ f\left({\boldsymbol h }\left(t\right),t\right) $来预测下一步的状态$ {\boldsymbol h }(t+\mathrm{\Delta }t) $. 更高级的求解方法,如DOPRI[45]会评估当前步骤的局部误差,并根据这个误差来自适应地调整步长$ \Delta  t $,以保证求解的准确性和稳定性.
常见的ODE数值求解方法总结如表2所示. 从形式上,根据未知数是否同时出现在等式两端可以将ODE数值求解方法分成显式求解法和隐式求解法. 其中,显式方法由于其简单高效,适用于非刚性问题;对于只在小时间间隔才有稳定解的方程,即刚性方程,隐式方法因其高稳定性而被优先选择.
此外,根据如何使用先前的信息来计算下一个值可以将ODE数值求解方法分成单步求解法和多步求解法. 单步求解法在每一步只依赖于当前步骤的信息来预测下一步的解. 单步方法通常更容易实现自适应时间步长机制,适合在计算精度和效率之间寻求平衡. 多步求解法在计算下一个值时会利用多个先前步骤的信息. 为了达到高阶精度,单步ODE求解器需要在每个时间步骤中更频繁地评估微分方程,因此将比多步ODE求解器产生更高的计算成本. 由于相对较低的成本,多步ODE求解器将会比单步ODE求解器更高效地计算出较准确的解. 然而,多步ODE求解器需要初始化计算,这种计算在每次重置计算或遇到不连续性时都会发生. 对于经常重置ODE求解器的应用来说,单步ODE求解器更高效.
本文根据静态图、离散动态图和连续动态图等不同类型图的特点,将相关工作分为静态图神经ODE、离散动态图神经ODE以及连续动态图神经ODE. 根据ODE所发挥的不同作用,将连续动态图神经ODE分为作为编码器和作为推断器2类. 根据图神经ODE的阶数将静态图神经ODE分为一阶静态图神经ODE和二阶静态图神经ODE. 此外,根据适用的场景不同,将离散动态图神经ODE分为时空图神经ODE和多智能体神经ODE. 本文提出的图神经ODE方法分类体系如图1所示. 本节中将对每类方法都归纳一个通式,再详细介绍相关的代表性工作.
图上的逐步消息传播方程可表示为:
$ {\boldsymbol H}(n+1)={\boldsymbol{AH}}\left(n\right)+{\boldsymbol H}\left(0\right). $
通过$ {\boldsymbol A}{\boldsymbol H}\left(n\right) $从其邻居那里学习节点信息,并通过$ {\boldsymbol H}\left(0\right) $记住其原始节点特征. 式(27)可得:
$ {\boldsymbol H}\left(n\right) =\left(\displaystyle\sum _{i=0}^{n}{\widehat{{\boldsymbol A}}}^{i}\right){\boldsymbol H}\left(0\right)={\left(\widehat{{\boldsymbol A}}-{\boldsymbol I}\right)}^{-1}\left({\widehat{{\boldsymbol A}}}^{n+1}-{\boldsymbol I}\right){\boldsymbol H}\left(0\right). $
$ {\boldsymbol H}\left(n\right) $结合了所有传播到$ n $步的信息和初始表示$ {\boldsymbol H}\left(0\right) $,能够实现在不忘记先前信息的情况下学习图结构.
为了进一步提高离散传播的有效性,Xhonneux等人[30]受现有基于图的扩散方法启发,提出了图神经ODE的开创性工作CGNN. 该方法引入一个连续变量$ \tau $,将$ {\boldsymbol H}\left(n\right) $视为从深度$ \tau =0 $到$ \tau =n $的积分的黎曼和1,自然地从离散传播过程过渡到连续情况. 之后,获取深度$ \tau $的表示$ {\boldsymbol H}\left(\tau \right) $的导数,并对其进行泰勒近似,将节点表示关于时间的导数定义为当前节点表示、邻居节点表示以及节点的初始值的组合,从而表征节点表示的连续动力学:
$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=\left(\widehat{{\boldsymbol A}}-{\boldsymbol I}\right){\boldsymbol H}\left(\tau \right)+{\boldsymbol H}\left(0\right),$
其中,$ \widehat{{\boldsymbol A}} $表示归一化邻接矩阵,即$ \widehat{{\boldsymbol A}} = \dfrac{\alpha }{2}({\boldsymbol I}+{{\boldsymbol D}}^{-1/2}{{\boldsymbol A}{\boldsymbol D}}^{-1/2}) $,$ \alpha $为超参数,用于控制扩散速率. 同时,为了允许不同特征通道之间的相互作用,引入权重矩阵${\boldsymbol W} $,针对多特征通道提出:
$ \dfrac{\mathrm{d}{\boldsymbol H}\left(\tau \right)}{\mathrm{d}\tau }=\left(\widehat{{\boldsymbol A}}-{\boldsymbol I}\right){\boldsymbol H}\left(\tau \right)+{\boldsymbol H}\left(\tau \right)\left({\boldsymbol W}-{\boldsymbol I}\right)+{\boldsymbol H}\left(0\right). $
该方法首次将神经ODE与GNN结合,有效地缓解了GNN的过平滑问题.
CGNN展现了静态图上图神经ODE的核心思想:将GNN和ODE系统结合起来,以数据驱动的方式学习复杂网络上的连续时间动力学. 首先,通过GNN提取出图上节点和边的特征,确保捕捉到节点间的互动及其邻近信息. 接下来,神经ODE模块以这些特征为初始值,通过$ n $阶ODE描述其随深度$ \tau $的演化动力学,则可以定义静态图上图神经ODE的通式,即:
$ \left\{\begin{aligned} & \dfrac{d^n\boldsymbol{H}\left(\tau\right)}{\rm{d}\tau^n}+\displaystyle\sum_{i=1}^na_i\left(\tau\right)\dfrac{d^i\boldsymbol{H}\left(\tau\right)}{\rm{d}\tau^i}=f\left(\tau\right), \\ & \boldsymbol{H}\left(0\right)=f_{\mathrm{E}}(\boldsymbol{A},\boldsymbol{X}),\end{aligned}\right. $
其中,初值$ {\boldsymbol H}\left(0\right) $为节点特征通过GNN等映射编码得到的嵌入,$ {f}_{\mathrm{E}}({\boldsymbol A},{\boldsymbol X}) $,$ {a}_{i}\left(\tau \right) $为各阶系数.
经过上述计算,图神经ODE最终输出特定深度上的节点、边或图的特征. 这些输出既可直接应用于预测任务,也可作为其他模块的输入. 但是一般情况下,一阶ODE能处理大部分的任务. 因此,许多工作聚焦于一阶图神经ODE的研究,即:
$ \left\{\begin{aligned} & \dfrac{\mathrm{d}\boldsymbol{H}\left(\tau\right)}{\mathrm{d}t}=f\left(\tau,\boldsymbol{H}\left(\tau\right),\boldsymbol{\theta}\right), \\ & \boldsymbol{H}\left(0\right)=f_{\mathrm{E}}(\boldsymbol{A},\boldsymbol{X}).\end{aligned}\right. $
静态图神经ODE代表性工作如表3所示.
1)NDE[58]. Hwang等人建立NDE模型模拟扩散过程及其固有不确定性. 该模型用一个编码层将节点特征矩阵转换为表示矩阵,采用离散的图拉普拉斯算子和全连接层建模网格网络上的扩散过程,从而建立了神经扩散ODE,即:
$ f\left(\boldsymbol{H}\left(\tau\right),\tau,\boldsymbol{\theta}\right)=\sigma\left(FC_2\left(\sigma\left(FC_1\left(\boldsymbol{H}\left(\tau\right)\right)\right)\right)\right), $
$ \dfrac{\mathrm{d}\boldsymbol{H}\left(\tau\right)}{\mathrm{d}\tau}=-\boldsymbol{K}\odot\boldsymbol{LH}\left(\tau\right)+f(\boldsymbol{H}\left(\tau\right),\tau;\boldsymbol{\theta}), $
其中$ \mathrm{F}\mathrm{C} $是全连接层,$ {\boldsymbol K}\odot {\boldsymbol L}{\boldsymbol H}\left(\tau \right) $为离散拉普拉斯算子,$ {\boldsymbol L} $是图拉普拉斯矩阵. NDE求解所得表示可用于单步和多步气候预测.
2)SANODE[26]. 增广神经ODE将增加的特征定义为具有固定大小和明确可解释性的节点特征梯度,能够学习高阶动力学. Zhang等人[26]利用这些性质,建立了二阶图神经ODE模型以提升GNN预测的可解释性. 首先,使用消息传递神经网络(message- passing neural networks,MPNN)更新节点表示为:
$ {{\boldsymbol h }}_{v}\left(\tau +1\right) ={f}_{\mathrm{U}}\left[{{\boldsymbol h }}_{v}\left(\tau \right),\displaystyle\sum _{u\in \mathcal{N}\left(v\right)}{f}_{\bf{M}}\left({{\boldsymbol h }}_{v}\left(\tau \right),{{\boldsymbol h }}_{u}\left(\tau \right)\right)\right],$
其中:
$ f_{\mathrm{U}}\left(\boldsymbol{x},\boldsymbol{y}\right)=\boldsymbol{x}+f_{\mathrm{G}}\left(\boldsymbol{x},\boldsymbol{y}\right), $
$ {f}_{\mathrm{G}}\left({\boldsymbol x},{\boldsymbol y}\right)={f}_{\mathrm{G}}\left(\displaystyle\sum _{u\in \mathcal{N}\left(v\right)}{f}_{\bf{M}}\left({\boldsymbol x},{\boldsymbol y}\right)\right). $
$ {f}_{\bf{M}} $计算邻居$ u $对节点$ v $传递的消息. 接着,将MPNN改写成差分方程的形式,得到一阶ODE:
$ \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau }={f}_{\mathrm{M}\mathrm{P}\mathrm{N}\mathrm{N}}^{\left(v\right)}\left({\boldsymbol H},{\boldsymbol \theta }\right)={f}_{\mathrm{G}}\left({{\boldsymbol h }}_{v}\left(\tau \right),{{\boldsymbol h }}_{\mathcal{N}\left(v\right)}\left(\tau \right)\right). $
之后,修改函数$ {f}_{\mathrm{U}} $的设定,推得二阶图神经ODE.
$ \dfrac{{d}^{2}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}{\tau }^{2}}+\alpha \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau } ={f}_{\mathrm{F}}\left({{\boldsymbol h }}_{v}\left(\tau \right)\right)+{f}_{\mathrm{G}}\left({{\boldsymbol h }}_{v}\left(\tau \right),{{\boldsymbol h }}_{\mathcal{N}\left(v\right)}\left(\tau \right)\right)\text{,} $
其中$ {f}_{\mathrm{F}}\left(\cdot \right) $代表力项,$ {f}_{\mathrm{G}}\left(\cdot \right) $代表阻尼项,该方法使用一阶ODE和二阶ODE共同预测,模拟物理系统,带来了直观的可解释性.
3)GraphCON[51]. Rusch等人[51]受基于物理微分方程的图机器学习模型启发,提出了二阶图神经ODE框架:
$ \dfrac{{d}^{2}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}{\tau }^{2}}+\alpha \dfrac{{\rm d}{{\boldsymbol h }}_{v}\left(\tau \right)}{{\rm d}\tau }=\sigma \left(f\left({\boldsymbol X},\tau ,{\boldsymbol \theta }\right)\right)-\gamma {\boldsymbol X}. $
通过耦合函数$ f $将GNN的传递机制与神经ODE结合. $ f $根据需要可部署为图注意力网络(GAT)或图卷积网络(graph convolutional network,GCN)[65]等.
综上,CGNN,NDE等一阶图神经ODE模型通过连续深度演化极大地缓解了GNN堆叠离散层引起的过平滑问题. 此外,Poli等人[55]还将带残差连接的GCN模型扩展为连续版本,提出了GCDE模型. SANODE,GraphCON等二阶ODE模型对连续深度演化进行更精细的建模,在应对过平滑和可解释问题上获得了更好的效果. 如表3所示,CGNN,SANODE和NDE基于消息传递或扩散建立图神经ODE,适用于疾病传播模型或信息传播分析等任务. GraphCON基于物理系统建立二阶图神经ODE,具有较好的可解释性. GCDE基于GCN建立图神经ODE,在一般图结构任务中具有较好的通用性和性能,能够处理多种复杂图结构任务.
在离散动态图中,神经ODE主要作为推断器,实现在已有快照的基础上推理出未来快照的状态. 给定图快照序列$ {\mathcal{G}}_{\mathrm{D}}=\{{G}^{1},{G}^{2},… ,{G}^{T}\} $,使用GNN处理每个时间点$ t $的图快照$ {G}^{t} $,得到该时间点的节点表示$ {\hat{{\boldsymbol H}}}^{t} $:
$ {\hat{{\boldsymbol H}}}^{t}={G}{N}{N}\left({{\boldsymbol A}}^{t},{{\boldsymbol X}}^{t}\right). $
之后,建立神经ODE用于模拟节点表示$ {{\boldsymbol H}}^{t} $随时间的连续变化. 该神经ODE可归纳为:
$ \dfrac{{\rm d}{{\boldsymbol H}}^{t}}{{\rm d}t}=f\left({{\boldsymbol H}}^{t-1},{\hat{{\boldsymbol H}}}^{t},t;{\boldsymbol \theta }\right)\text{,} $
其中$ {\hat{{\boldsymbol H}}}^{t} $是由非连续处理得到的当前时间点$ t $的节点表示,而$ {{\boldsymbol H}}^{\mathit{t}} $代表连续更新后的节点表示. 使用数值求解方法来求解上述ODE,可得到任意时间点的连续节点表示$ {{\boldsymbol H}}^{t} $用于预测未观测到的节点或边状态,进而解决下游任务. 离散动态图神经ODE代表性工作如表4所示.
1)NDCN. 现实世界网络中节点的互动结构复杂,控制复杂网络中节点状态动态变化的规则是连续时间且非线性的,导致系统内的结构和动态依赖关系难以通过显式数学函数建模. 为解决这些问题,Zang等人[31]提出了NDCN模型,将神经ODE和GNN结合起来学习图上的连续时间动力学. NDCN模型概括为:
$ \begin{array}{c}\underset{{\boldsymbol W}\left(t\right),{\boldsymbol \varTheta }\left(T\right)}{\mathrm{arg\ min}}\mathcal{L}=\displaystyle\int _{0}^{T}{f}_{\mathcal{R}}\left({\boldsymbol H},G,{\boldsymbol W},t\right)\mathrm{d}t+{f}_{\mathcal{S}}\left.\left(Y\left({\boldsymbol H}\left(T\right)\right);{\boldsymbol \varTheta }\right)\right),\end{array} $
$ \rm{s.t.}\left\{\begin{aligned} & \boldsymbol{H}'^t=f_{\mathrm{E}}\left(\boldsymbol{H}^{\mathit{t}},\boldsymbol{W}_{\mathrm{E}}\right), \\ & \dfrac{\rm{d}\boldsymbol{H}'^t}{\rm{d}t}=f\left(\boldsymbol{H}',G,\boldsymbol{W}_{\mathrm{h}},t\right), \\ & \boldsymbol{H}^{\mathit{t}}=f_{\mathrm{D}}\left(\boldsymbol{H}'^t,\boldsymbol{W}_{\mathrm{D}}\right).\end{aligned}\right. $
在NDCN框架中,$ {f}_{\mathrm{E}} $为编码函数,$ {f}_{\mathrm{D}} $为解码函数. $ \displaystyle\int _{0}^{T}{f}_{\mathcal{R}}\left({\boldsymbol H},G,{\boldsymbol W},t\right)\mathrm{d}t $是从$ t=0 $~T 期间上连续时间动力学的“运行”损失,而$ {f}_{\mathcal{S}}\left(Y\left({\boldsymbol X}\left(T\right)\right),{\boldsymbol \varTheta }\right) $是在时间$ T $时刻的“终端”损失,$ Y $表示采用独热编码的节点语义标签.
在实际部署中,损失函数和编/解码器的选择依照任务需求来确定. 例如在模拟网络连续动力学时,Zang等人[31]采用$ {L}_{1} $-范数损失作为运行损失,采用二层全连接网络、单层GCN和线性层分别作为ODE函数中的$ {f}_{\mathrm{E}} $、$ f $和$ {f}_{\mathrm{D}} $来模拟图上的瞬时动态变化,即:
$ \begin{aligned}&\underset{{{\boldsymbol W}}_{\mathit{*}},{{\boldsymbol b }}_{\mathit{*}}}{\mathrm{arg\ min}}\mathcal{L}=\displaystyle\int _{0}^{T}\left|{{\boldsymbol H}}^{t}-{\hat{{\boldsymbol H}}}^{t}\right|\mathrm{d}t,\\ &\mathrm{s}. \mathrm{t}\left\{\begin{aligned}& {\boldsymbol H'^{t}}=\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left({{\boldsymbol H}}^{\mathit{t}}{{\boldsymbol W}}_{\mathrm{E}}+{{\boldsymbol \beta }}_{\mathrm{E}}\right){{\boldsymbol W}}_{0}+{{\boldsymbol \beta }}_{0},\\ & \dfrac{{\rm d}{{\boldsymbol H'}}^{t}}{{\rm d}t}=\mathrm{G}\mathrm{C}\mathrm{N}\left({\boldsymbol H'},{\boldsymbol A}\right),\\ & {{\boldsymbol H}}^{\mathit{t}}={{\boldsymbol H'}}^{t}{{\boldsymbol W}}_{\mathrm{D}}+{{\boldsymbol \beta }}_{\mathrm{D}}, \end{aligned}\right.\end{aligned} $
其中$ {{\boldsymbol W}}_{\mathrm{*}} $和$ {{\boldsymbol \beta }}_{\mathrm{*}} $为待优化的参数,$ {\hat{{\boldsymbol H}}}^{t} $为现实动力学监督信号.
2)STG-ODE[46]. 现有的工作大部分结合GNN和RNN来分别获得空间表示和时间表示. 但是,一方面,浅层的GNN不能捕获长距离的空间相关性,而过平滑现象使得GNN的深度受到限制,这导致节点之间的长距离依赖往往被忽略;另一方面,这些工作只考虑了空间连接,忽略了大量的语义连接,影响了对交通网数据的全面理解. Fang等人[46]基于空间连接性和交通流的语义相似性,构建了空间邻接矩阵和语义邻接矩阵,来考虑节点间的物理距离和上下文相似性,通过基于张量的ODE来捕捉时空动力学,即:
$ {\mathcal{H}}^{(l+1)}={\mathcal{H}}^{\left(l\right)}{\times }_{1}\hat{{\boldsymbol A}}{\times }_{2}{\boldsymbol U}{\times }_{3}{\boldsymbol W}+{\mathcal{H}}^{\left(0\right)},$
其中,$ {\mathcal{H}}^{\left(l\right)}\in {\mathbb{R}}^{N\times T\times F} $是一个空间-时间张量,表示第$ l $层节点的隐向量,$ {\boldsymbol U} $是时序转换矩阵,${\boldsymbol W} $是特征转换矩阵,$ {\times }_{i} $表示在模$ i $上的张量-矩阵乘法. 其使用与CGNN是有相似的理论推导,将离散公式扩展到连续表达式2,即:
$ \begin{split} \dfrac{\mathrm{d}\mathcal{H}\left(\tau \right)}{\mathrm{d}\tau }=&\mathcal{H}\left(\tau \right){\times }_{1}\left(\hat{{\boldsymbol A}}-{\boldsymbol I}\right)+\mathcal{H}\left(\tau \right){\times }_{2}\left({\boldsymbol U}-{\boldsymbol I}\right)+\\&\mathcal{H}\left(\tau \right){\times }_{3}\left({\boldsymbol W}-{\boldsymbol I}\right)+\mathcal{H}\left(0\right).\end{split} $
3)MTGODE[47]. 该方法被用于应对多元时间序列建模中离散架构导致的不连续的潜在状态轨迹,以及高预测误差、高计算开销和依赖图先验等挑战. 该方法首先将多元时间序列抽象为具有时间演化节点特征和未知图结构的动态图. 然后,设计并求解一个神经ODE来补充缺失的图拓扑,并统一空间和时间的信息传递,允许更深层次的图传播和细粒度的时间信息聚合,以表征稳定且精确的空间-时间动力学. 即:
$ \begin{split} \dfrac{\mathrm{d}{{\boldsymbol H}}_{{T}}^{t}}{\mathrm{d}t}=&\mathcal{P}\Bigg(\mathcal{A}\Bigg({{O}{D}{E}{S}{o}{l}{v}{e}}^{2}\Bigg({T}{C}{N}\left({{\boldsymbol H}}_{{T}}^{t},t,{\boldsymbol \theta }\right),\\& \dfrac{\mathrm{d}{{\boldsymbol H}}_{{{\mathrm{G}}}}^{\tau }}{\mathrm{d}\tau },0,… ,T\Bigg),{\boldsymbol \varPhi }\Bigg),R\Bigg)\text{,} \end{split}$
其中$ R $表示感受野,$ \mathrm{T}\mathrm{C}\mathrm{N} $为时序卷积网络. $ \mathcal{P}(\cdot ) $表示填充函数. 通过将由$ \dfrac{\mathrm{d}{{\boldsymbol H}}_{\mathrm{G}}^{\tau }}{\mathrm{d}\tau } $定义的空间ODE视为由$ \dfrac{\mathrm{d}{{\boldsymbol H}}_{\mathrm{T}}^{t}}{\mathrm{d}t} $定义的时间神经ODE的内部过程,允许模型以完全连续的表征输入多元时间序列底层交织的空间-时间动力学,为下游任务提供更具表现力的表示.
4)CG-ODE[48]. 许多现实世界系统的动态特性不仅表现在节点特征的演化上,而且也表现为图结构的动态变化. Huang等人[48]考虑了节点和边的共同演化,建立了耦合图神经ODE. 该模型首先通过时空GNN处理构建的时序图,得到综合了历史观测信息和结构信息的节点表示. 之后,采用自注意力机制为每个对象生成序列表示,然后利用这种表示来计算潜在节点初始状态的后验分布,并从中抽样得到节点的潜在初始状态. 接着,通过函数$ {f}_{\text{edge}} $编码边的端点表示得到边的潜在初始状态$ {{\boldsymbol b }}_{ij}^{0} $. 而后,通过考虑节点和边的动态关系以及它们随时间的自我演化,采用由耦合的ODE函数定义的生成模型来计算未来预测的节点和边的潜在状态,即:
$ \dfrac{{\rm d}{{\boldsymbol H}}^{t}}{{\rm d}t}=\sigma \left({{\widetilde{\boldsymbol A}}}^{t}{{\boldsymbol H}}^{t}{\boldsymbol W}\right)-{{\boldsymbol H}}^{t}+{{\boldsymbol H}}^{0}, $
$ \begin{array}{c}\dfrac{{\rm d}{{\boldsymbol b }}_{ij}^{t}}{{\rm d}t}={f}_{\mathrm{e}}\left([{{\boldsymbol h }}_{{v}_{i}}^{t}\|{{\boldsymbol h }}_{{v}_{j}}^{t}]\right)+{f}_{\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{f}}\left({{\boldsymbol b }}_{ij}^{t}\right),\end{array} $
$ {{\boldsymbol A}}_{i,j}^{t}={f}_{\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{e}2\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}}\left({{\boldsymbol b }}_{ij}^{t}\right),\;\;{{\widetilde{\boldsymbol A}}}^{t}={\left({{\boldsymbol D}}^{t}\right)}^{-1}{{\boldsymbol A}}^{t},$
其中,$ {f}_{\mathrm{e}} $为将2个节点的串联转换为它们对应边的潜在状态的映射函数,$ {f}_{\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{f}} $为负责边的自我进化的函数. $ {f}_{\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{e}2\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}} $将潜在边状态转化为标量,$ {{\widetilde{\boldsymbol A}}}^{t} $计算来自邻居的消息传递,并随边状态的改变而实时更新.
随后运行解码器,将每个解码分布的均值视为边和节点的预测值. 其中,节点解码函数根据节点的潜在状态预测每个节点在不同时间点的状态,而边解码函数则根据边的潜在状态预测对象间的交互强度.
5)HOPE[61]. 基于空间的图神经网络会错过图谱中嵌入的高阶非邻域语义,这将导致无法提供潜在状态表示的适当初始化. 此外,生成型图ODE模型通常只涉及一阶导数,无法捕捉潜在表示中的高阶依赖. 因此,为了解决上述问题,Luo等人[61]巧妙结合了二阶图卷积以捕获非邻域语义信息,以及利用二阶图ODE来模拟更高阶的时间依赖性. 具体来说,该方法首先构建一个离散动态图,然后使用2个并行的GNN分支从互补的视图中提取整体时空关系. 一方面,采用自注意机制来自适应地聚合邻域信息,从空间视角学习对象表示. 另一方面,采用二阶谱图卷积,利用图谱中嵌入的语义学习非邻域节点之间的高阶相关性. 之后,将这些时间节点表示聚合成顺序表示,以初始化每个节点和边的潜在状态. 然后,使用二阶ODE模拟节点和边的演化. 其中,通过节点ODE聚合邻域信息,并随着时间的推移迭代更新状态表示:
$ \dfrac{{\mathrm{d}}^{2}{{\boldsymbol H}}^{t}}{\mathrm{d}{t}^{2}}+\gamma \dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}=\sigma \left({\left({\hat{{\boldsymbol D}}}^{t}\right)}^{-1}{{{{\widehat{\boldsymbol A}}_{\mathrm{U}}}}^{\mathit{t}}{\boldsymbol H}}^{t}{{\boldsymbol W}}_{{\mathrm{p}}}\right)-{{\boldsymbol H}}^{t}, $
其中可不断更新的邻接矩阵$ {\widehat{\boldsymbol A}}_{\mathrm{U}} $同CG-ODE模型相似,由边ODE建模边潜在状态表示的演化得到,即:
$ \dfrac{d^2\boldsymbol{b}_{ij}^t}{d^2t}+\gamma\dfrac{\rm{d}\boldsymbol{b}_{ij}^t}{\rm{d}t}=\rho_{\mathrm{n}}\left([\boldsymbol{h}_{v_i}^t\parallel\boldsymbol{h}_{v_j}^t\parallel\boldsymbol{h}_{v_i}^t\odot\boldsymbol{h}_{v_j}^t]\right)+\rho_{\mathrm{e}}\left(\boldsymbol{b}_{ij}^t\right), $
$ {{\widetilde{\boldsymbol A}}}_{ij}^{t}={\rho }_{\mathrm{s}}\left({{\boldsymbol b }}_{ij}\right),$
其中$ {\rho }_{\mathrm{e}} $和$ {\rho }_{\mathrm{n}} $为将输入空间转换到表示空间的函数,$ {\rho }_{\mathrm{s}} $为将边状态转换成标量的映射. 最后,基于潜在状态表示输出节点和边的预测值.
6)SEGODE[27]. SEGODE框架通过结合图表示学习和神经ODE,有效地捕获时间图的动态演化规律. 该模型分为3个主要模块:结构编码、演化学习和表示解码. 首先,结构编码模块利用邻接矩阵、GDV(graphlet degree vector)矩阵和PageRank矩阵提取每个网络快照的局部、高阶和全局结构信息,并通过相应的编码器获取结构表示. 之后,从结构编码中获得的离散时间嵌入通过插值投影转化为连续时间表示$ {{\boldsymbol H}}_{{\mathrm{ip}}}^{t} $.
接着,在演化学习模块中,应用注意力机制创建的ODE函数对结构嵌入进行时间更新. 即:
$ \dfrac{\rm{d}\boldsymbol{H}^t}{\rm{d}t}=softmax\left(\dfrac{\boldsymbol{Q}^t\boldsymbol{K}^t}{\sqrt{d_k}}\right)^{\mathrm{T}}\boldsymbol{V}^t, $
$ \boldsymbol{Q}^t=\boldsymbol{H}_{\mathrm{ip}}^t\boldsymbol{W}_{\mathit{q}},\; \; \boldsymbol{K}^t=\boldsymbol{H}_{\mathrm{ip}}^t\boldsymbol{W}_{\mathit{k}},\; \; \boldsymbol{V}^t=\boldsymbol{H}^t, $
其中,$ {{\boldsymbol W}}_{\mathit{q}} $,$ {{\boldsymbol W}}_{\mathit{k}} $为权重矩阵. 随后利用ODE求解器更新节点隐藏嵌入,学习网络随时间的演化.
最后,解码模块通过2个全连接层生成节点的最终表示,这些表示经过优化后用于链路预测任务.
综上,离散动态图ODE作为推断器的核心思想是利用ODE的连续时间建模能力来桥接离散时间点之间图快照的状态变化,从而能够更加精确和连续地推理和预测图快照的状态演化. NDCN是首个建立了兼顾网络动力学和结构序列预测的框架. 由表4可见,宏观上,STG-ODE在构建ODE函数时通过张量同时考虑了时间和空间. ASTGCN构建ODE函数时同时考虑了局部和全局. 但是,STG-ODE,ASTGCN等模型只能实现在空间上连续聚合,在时序演化上仍然离散. 而MTGODE将空间ODE视为其所提出的时间神经ODE的内部过程,实现了时域和空域都连续、允许模拟完全连续的潜在时空动力学. 微观上,CG-ODE和HOPE等进一步关注于图结构的动态变化,在建立ODE方程时考虑了节点和边的共同演化,更加准确地捕获图结构的动态变化. 此外,HOPE和SEG-ODE模型还引入了高阶依赖性的捕捉,将高阶导数与高阶GNN结构编码机制结合,增强了模型对复杂网络动力学的理解和表示. SEG-ODE模型的自注意力机制和GNG-ODE的门控机制更灵活地捕捉节点间的重要历史信息. 与上述模型不同,MTGODE不依赖于图先验,而通过神经ODE来补充缺失的图拓扑.
在连续动态图中,图数据的观测值可能会在不均匀的时间点上发生,离散动态图上基于固定时间间隔的模型难以适应这种情况. 但是,神经ODE可以自然地处理不均匀的时间戳,将ODE作为编码器可以在任何观测到的时间点上直接进行状态的评估和预测. 此外,连续动态图还存在信息缺失的问题,而神经ODE作为推断器时通过在连续时间域内建模,能够在没有直接观测数据的时间点上进行预测,增强了模型的泛化能力. 因此,将连续动态图与神经ODE结合成为重要的研究方向. 连续动态图ODE代表性工作如表5所示.
图神经ODE作为编码器用于连续动态图的目标是捕获和学习节点表示的连续动态. 其遵循以下框架:
首先,对每个事件$ \varepsilon $生成消息$ {m}_{\mathrm{\varepsilon }} $. 对于节点$ v $,聚合所有时间邻域内相关事件生成的消息. 接着,定义ODE模型来描述节点状态随时间的连续变化. 即:
$ \dfrac{{\rm d}{{\boldsymbol h }}_{v}^{t}}{{\rm d}t}=f\left({{\boldsymbol h }}_{v}^{t},{A}{G}{G}{R}{E}{G}{A}{T}{E}\left(\left\{{m}_{\varepsilon }\right|\varepsilon =\left(u,v,{t}\right)\in {\mathcal{E}}_{\mathcal{G}}\}\right)\right). $
在事件发生时,利用ODE模型和数值求解方法更新节点状态. 经过一系列事件的处理和节点状态的连续更新后,最终每个节点在任意时间点$ t $的状态表示$ {{\boldsymbol h }}_{v}^{t} $用于下游任务.
1)ConTIG. 现有的动态嵌入方法在交互图上仅在交互发生时才离散地更新节点嵌入,它们未能捕捉节点嵌入轨迹的连续动态演化,因此失去了在长间隔交互中不可忽视的非活跃节点信息. Wang等人[32]提出的ConTIG模型,旨在学习节点嵌入轨迹并捕捉节点表示的连续动力学. 当一个交互信息到来时,首先将交互的时间戳投影到连续时间空间中,生成其时间嵌入. 使用一个全连接层作为编码器将输入交互节点的最新交互信息投影到一个潜在空间中,即:
$ {F}{C}\left({{\boldsymbol h }}_{v}^{{t}_{p-1}}\| {{\boldsymbol h }}_{u}^{{t}_{p-1}}\| {{\boldsymbol l}}_{vu}\| {F}_{T}\left(\Delta t\right)\right),u\in \mathcal{N}\left(v\right), $
其中,$ {t}_{p} $为时间戳,$ {{\boldsymbol h }}_{v}^{{t}_{p-1}} $为节点$ v $在$ {t}_{p} $前最新的表示,$ {{\boldsymbol h }}_{u}^{{t}_{p-1}} $为节点$ u $在$ {t}_{p} $前最新的表示,$ {{\boldsymbol l}}_{vu} $为最新的交互信息的表示,$ {F}_{T}\left(\Delta t\right) $是时间$ \Delta t $的嵌入.
之后,将得到的表示作为定义的神经ODE的初始值. 式(58)整合了最新交互、邻居特征和固有特征,即:
$ \dfrac{\rm{d}\boldsymbol{H}^{t_p}\left(\tau\right)}{\mathrm{d}\tau}=z_l\odot\boldsymbol{H}^{t_p}\left(0\right)+z_n\odot\boldsymbol{A}_p\boldsymbol{H}^{t_p}\left(\tau\right)-z_i\odot\boldsymbol{H}^{t_p}\left(\tau\right), $
其中,$ {{\boldsymbol A}}_{p} $为$ {t}_{p} $时的归一化邻接矩阵. $ {{\boldsymbol H}}^{{t}_{p}}\left(0\right) $考虑到了最新的交互信息可能对当前交互产生重大影响,$ {{\boldsymbol A}}_{p}{{\boldsymbol H}}^{{t}_{p}}\left(\tau \right) $考虑到节点状态还受到随时间推移其邻居注意力的影响,$ {{\boldsymbol H}}^{{t}_{p}}\left(\tau \right) $考虑到节点的固有特征将决定状态,能够成功捕获节点嵌入轨迹的连续动力学. $ {{z}_{l},z}_{n},{z}_{i} $是3个门控机制,用于自适应地结合最新交互、邻居特征以及固有特征这3个因素. 使用神经ODE结束时的节点嵌入来更新交互节点的嵌入.
2)NeurTWs[62]. 为了应对连续动态图的不规则交互,与先前的工作在聚合邻域信息时将节点属性与额外的时间编码串联起来的做法不同,NeurTWs显式地建模时间依赖性. 首先,通过时序随机游走和匿名化处理近似动态图Motif提取,得到匿名时序游走$ \widehat W $. 针对每个动态图Motif,模型先对一系列不同时间点的节点执行连续演化操作,整合多个交互时间间隔,利用ODE显式模拟节点之间潜在的空间时间动力学,即:
$ \dfrac{\mathrm{d}\boldsymbol{H}_i}{\mathrm{d}t}=\boldsymbol{H}_{i-1}+\displaystyle\int_{t_{i-1}}^{t_i}f\left(\boldsymbol{H}_t,\boldsymbol{\theta}\right)\rm{d}t, $
$ A'\left(w_i\right)=MLP\left(A\left(w_i\right);\boldsymbol{\psi}\right), $
$ {{\boldsymbol H}}_{i}={R}{N}{N}\left(\dfrac{\mathrm{d}{{\boldsymbol H}}_{i}}{\mathrm{d}t},{A'}\left({w}_{i}\right);{\boldsymbol \phi }\right)\text{,} $
其中,$ {w}_{i} $表示时序游走$ \mathcal W $中的节点,$ A\left({w}_{i}\right) $表示匿名化操作. $ {{\boldsymbol H}}_{i-1} $为编码$ \left(A\left({w}_{i-1}\right),{t}_{i-1}\right)\in \widehat{W} $后得到的潜在状态,$ {A'}\left({w}_{i}\right) $为对离散观测$ A\left({w}_{i}\right) $的线性映射,$ {\boldsymbol \psi } $和$ {\boldsymbol \phi } $为神经网络参数. 随后,模型对连续演化的结果进行调整,利用瞬时输入激活潜在状态轨迹,确保模型能够准确反映每个时间点的具体信息.
综上,图神经ODE作为连续动态图编码器的核心思想是将图的动态变化和节点之间基于事件的复杂交互编码为一个连续的状态空间模型. 由表5可见,在设计ODE函数时,ConTIG,GDERec等模型侧重使用图结构来增强ODE的表达能力,通过使用门控机制、二阶导数等使模型能够在连续的时间框架内更有效处理结构化数据. NeurTWs,GCDE-RNN,GNODEe等模型更新依赖于前一个状态,结合了递归神经网络来处理图结构数据中的时间依赖性,适合对图数据的结构和时间信息进行同时学习. 在消息聚合时,ConTIG通过全连接层处理不同节点的特征以及时序信息,强调在处理节点特征时融入时序信息,适用于节点状态随时间显著变化的场景. GDERec模型使用哈达玛积结合来自不同时间步的节点状态,增强了特征的非线性表达能力. NeurTWs,GCDE-RNN,GNODEe则采用循环神经网络结构,帮助模型捕捉长期的时间依赖关系,能够更好地适应和解析图数据的复杂性和动态性.
将神经网络应用于不规则采样的数据时,先前的工作通常是将观测结果放入固定持续时间的区间内,而潜在动力学也以同样的方式离散化. 这导致了数据缺失和不明确的潜变量的困难. 为了应对上述挑战,连续动态图ODE作为推断器构建潜在状态在所有时间点上都被定义的连续时间模型. 其满足通式:
$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t}}{{\rm d}t}=f\left({{\boldsymbol h }}_{{v}_{1}}^{t},{{\boldsymbol h }}_{{v}_{2}}^{t},… ,{{\boldsymbol h }}_{{v}_{N}}^{t};{\boldsymbol \theta }\right)\text{,} $
其中,$ {v}_{1},{v}_{2},… ,{v}_{N} $等节点在时间戳$ t $时的状态$ {{\boldsymbol h }}_{{v}_{1}}^{t},{{\boldsymbol h }}_{{v}_{2}}^{t},… , {{\boldsymbol h }}_{{v}_{N}}^{t} $由这些节点时间$ t $之前的状态计算得到,其实现了从已知的初始状态出发来精确地推断图的连续演化.
1)LG-ODE[56]. 为了处理多智能体系统的不完全观测,Huang等人[56]建立了基于变分自编码器的潜变量ODE模型,该模型由编码器、ODE参数化的生成模型、解码器3部分组成,并进行联合训练. 首先,编码器进行动态节点表示学习,学习一个编码函数,为每个观测输出结构化的上下文表示. 即:
$ {\boldsymbol{z}}_{i}^{t}={f}_{\mathrm{E}}\left({v}_{i},\left.\{{v}_{j}\right|\mathrm{i}\mathrm{f}\ {v}_{j}\in \mathcal{N}\left({v}_{i}\right)\}\right) \text{.} $
然后,通过时间自注意力机制将结构化的观测表示聚合成每个对象固定维度的序列表示$ {{\boldsymbol u}}_{i} $,即:
$ {{\boldsymbol u}}_{i}={f}_{\text{аggre}}\left({{\boldsymbol z}}_{i}^{{t}_{1}},{{\boldsymbol z}}_{i}^{{t}_{2}},…,{{\boldsymbol z}}_{i}^{{t}_{{T}_{i}}}\right). $
接着,模型从编码器计算的近似后验分布中采样得到每个对象的潜在初始状态,即:
$ q_{\boldsymbol{\phi}}\left(\boldsymbol{h}_{v_i}^0|v_1,v_2,\dots,v_N\right)=Gaussian\left(\boldsymbol{\mu}_{\boldsymbol{h}_{v_i}^0},\boldsymbol{\sigma}_{\boldsymbol{h}_{v_i}^0}\right), $
$ {{\boldsymbol \mu }}_{{{\boldsymbol h }}_{{v}_{i}}^{0}},{{\boldsymbol \sigma }}_{{{\boldsymbol h }}_{{v}_{i}}^{0}}={f}_{\mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}}\left({{\boldsymbol u}}_{i}\right)\text{,} $
其中,$ {{\boldsymbol \mu }}_{{{\boldsymbol h }}_{{v}_{i}}^{0}},{{\boldsymbol \sigma }}_{{{\boldsymbol h }}_{{v}_{i}}^{0}} $为从序列表示$ {{\boldsymbol u}}_{i} $计算得到的高斯分布的均值和方差.
之后,使用GNN作为ODE函数来模拟对象之间的连续交互,并且通过求解ODE得到从初始状态到最终时间$ T $状态的序列. 即:
$ \dfrac{\rm{d}\boldsymbol{h}_{v_i}^t}{\rm{d}t} = f\left(\boldsymbol{h}_{v_1}^t,\boldsymbol{h}_{v_2}^t,\dots,\boldsymbol{h}_{v_N}^t\right) = f_O\left(\displaystyle\sum_{j\in\mathcal{N}_i}^{ }f_R\left([\boldsymbol{h}_{v_i}^t,\boldsymbol{h}_{v_j}^t]\right)\right), $
$ {{\boldsymbol h }}_{{v}_{i}}^{0},{{\boldsymbol h }}_{{v}_{i}}^{1}, …, {{\boldsymbol h }}_{{v}_{i}}^{T} = {O}{D}{E}{S}{o}{l}{v}{e} \left(f, \left[{{\boldsymbol h }}_{{v}_{1}}^{0},{{\boldsymbol h }}_{{v}_{2}}^{0}, …,{{\boldsymbol h }}_{{v}_{N}}^{0}\right] , \left({t}_{0},{t}_{1}, …,{t}_{T}\right) \right)\text{,} $
其中,函数$ {f}_{O} $处理邻域信息,$ {f}_{R} $为关系映射函数.
2)TANGO. Huang等人[52]利用图神经ODE框架来建模多智能体动态系统,并引入时间反演对称性,以增强模型对系统能量变化的控制和对非保守系统的适应性. 首先,将多智能体系统表示为图并使用基于Transformer的时空GNN进行编码. 之后,采用GNN作为ODE函数,驱动系统状态向前演化,输出每个连续时间$ t $和每个智能体$ i $的潜在状态. 即:
$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}}{{\rm d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({{\boldsymbol h }}_{{v}_{1}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}},{{\boldsymbol h }}_{{v}_{2}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}},…,{{\boldsymbol h }}_{{v}_{N}}^{t,\mathrm{f}\mathrm{w}\mathrm{d}}\right). $
确定前向轨迹在最终时间点$ t $的状态后将其作为反向轨迹的起始点. 相似地,采用ODE函数描述系统状态随时间的变化规律,模拟时间的反向流动. 随后,利用ODE求解器沿时间反方向积分求解,逐步构建出反向轨迹,实现了对系统前向演化过程的时间反演,即:
$ \dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}}}{{\rm d}t}={f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({{\boldsymbol h }}_{{v}_{1}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}},{{\boldsymbol h }}_{{v}_{2}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}},…, {{\boldsymbol h }}_{{v}_{N}}^{t,\mathrm{r}\mathrm{e}\mathrm{v}}\right). $
接着,设计时间反演对称性损失,通过最小化模型预测的前向轨迹与反向轨迹之间的差异,保证模型遵循物理规律. 该模型不仅能够准确预测多智能体系统的动力学,还能够通过时间反演对称性损失有效地捕捉系统动力学的基本物理原则,从而为动态系统建模提供了一种更为通用和鲁棒的解决方案.
3)GG-ODE. 为了应对离散模型在跨环境学习的连续多智能体系统长期预测方面精度较低的挑战,Huang等人[57]将多智能体图神经ODE模型拓展到跨环境学习的连续多智能体系统动力学的研究中. 该方法假设不同环境下的动力学受到共享ODE函数捕获的共同物理定律的支配,从而实现模型泛化,即:
$ \begin{array}{c}\dfrac{{\rm d}{{\boldsymbol h }}_{{v}_{i}}^{t,{\mathrm{e}}}}{{\rm d}t}=\displaystyle\sum _{j\in {\mathcal{N}}_{i}}{f}_{\mathrm{G}\mathrm{N}\mathrm{N}}\left({\widehat{{\boldsymbol h }}}_{{v}_{i}}^{t,{\mathrm{e}}},{\widehat{{\boldsymbol h }}}_{{v}_{j}}^{t,{\mathrm{e}}}\right)+{f}_{\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{f}}\left({\widehat{{\boldsymbol h }}}_{{v}_{i}}^{t,{\mathrm{e}}}\right)\text{,}\end{array} $
$ {\widehat{{\boldsymbol h }}}_{{v}_{i}}^{t,{\mathrm{e}}}={f}_{\mathrm{e}\mathrm{n}\mathrm{v}}\left({{\boldsymbol h }}_{{v}_{i}}^{t,{\mathrm{e}}}\|{{\boldsymbol u}}^{{\mathrm{e}}}\right)\text{,} $
其中$ {{\boldsymbol u}}^{{\mathrm{e}}} $表示环境编码器$ {f}_{\mathrm{e}\mathrm{n}\mathrm{v}} $从观测轨迹中自动提取外源因素(如温度、重力等)的潜在影响,输出这些因素的表示外源因素的潜在表示.
此外,由于该模型考虑外源因素的影响,额外设计了2个正则化损失:通过互信息最小化强制学习的初始状态与外源因素之间的正交性;以及通过对比学习减少同一系统内学习的外源因素的时间方差. 实现了在长期范围内,能够很好地泛化到观测数据较少的新系统.
综上,图神经ODE作为连续动态图编码器的核心思想是构建一个连续时间动态系统来模拟图中不规则采样得到的节点状态随时间的演化. LGODE模型以基于事件的编码器得到的表示为ODE模型的初始值,建立由ODE参数化的生成模型,学习推理复杂系统的交互和动力学. TANGO,GG-ODE等模型则引入时间反演对称性、反事实估计等进一步扩展了对多智能体系统动力学的建模. 由表5可见,LGODE强调节点的动态表示和序列表示学习,适合于高时间分辨率和长短期依赖关系强的应用场景. TANGO通过编码函数结合图和时间窗信息进行聚合,适合于需要在特定时间窗内捕捉节点动态变化的场景. GGODE通过结合环境聚合信息和节点当前状态来进行更新,适合于需要综合考虑节点和其环境变化的应用.
将动态图视为时间上离散的图快照不足以捕捉本质上是连续的底层动力学. 为了克服这种缺陷,Gao等人[64]考虑把图结构在全连接的神经ODE基础上进行扩展,提出了GNODE模型. GNODE模型的编码器GNODEe采纳了与ODE-RNN相似的算法,是由GNODE整合到GRNN[66]得到的增广图序列编码器,即:
$ {{\boldsymbol H}}^{{t}_{i+1}}={{G}{N}{O}{D}{E}}^{\mathrm{e}}\left({G}{R}{N}{N}\right({{\boldsymbol H}}^{{t}_{i}}{,G}^{i}),\left\{{t}_{i},{t}_{i+1}\right\}) \text{,} $
$ {G}{R}{N}{N}(\cdot ) $为堆叠的图循环神经网络,其策略是先通过GCN对图进行编码,将其转换成潜在状态,然后使用RNN处理这些潜在状态的序列. $ {G}{R}{N}{N}(\cdot ) $也可采用编码和序列处理都为GCN的策略.
编码器GNODEe首先将连续观测序列$ \{{G}^{i},{t}_{i}|i= \mathrm{1,2},… \} $编码成潜在状态$ {{\boldsymbol H}}^{{t}_{I}} $,并通过参数化潜在状态变分分布,从中采样出未来的初始潜在状态. 采样的初始潜在状态被输入到推断器GNODEi中,以预测潜在状态的未来演化. 即:
$ \boldsymbol{H}^{t'_{0'}}\sim q\left(\boldsymbol{H}^{t'_{0'}}|GNODE^{\mathrm{e}}\left(\left\{G^i,t_i|i=\mathrm{1,2,}\dots,I\right\}\right)\right), $
$ {{\boldsymbol H}}^{{t'}}={{G}{N}{O}{D}{E}}^{\mathrm{i}}({{\boldsymbol H}}^{t'_{0'}},{t'}). $
然后使用一个概率图解码器从条件分布$ p({G}^{{i'}}|{\boldsymbol H}\left({\mathit{t'}}_{ {\mathit{i'}}}\right)) $生成图,即:
$ {\widehat{G}}^{{i'}}\sim p\left({G}^{{i'}}|{{\boldsymbol H}}^{t'_{i'}}\right)=p\left({G}^{{i'}}|{{\boldsymbol H}}^{{t'}}\right)\text{,} $
其中,推断器GNODEi通过可学习的GCN来参数化:
$ \dfrac{\mathrm{d}{{\boldsymbol H}}^{t}}{\mathrm{d}t}={G}{C}{N}({{\boldsymbol H}}^{t},{\widetilde{D}}_{g}({{\boldsymbol H}}^{t})). $
该方法没有采用手动选择的时间戳和分段固定的邻接矩阵,而是用图解码器的输出替换GCN中输入的邻接矩阵A. 其中,$ {\widetilde{D}}_{g}(\cdot ) $为Gumbel-Softmax层.
本节将对图神经ODE研究相关和常见的实验任务进行总结,包括节点分类、链接预测、序列预测和轨迹预测. 具体地,将介绍每种任务的操作步骤、评价指标和对应的数据集. 基线方法和数据集的详细信息已经集成在一个可访问的公开网站3上,以便于对这一快速发展主题进行开放研究.
对于给定的静态图$ G=(\mathcal{V},\mathcal{E},\mathcal{X}) $,$ {\mathcal{V}}_{{\mathrm{l}}} $表示有标签的节点集,且$ {\mathcal{V}}_{{\mathrm{l}}}\subset \mathcal{V} $,$ {\mathcal{V}}_{{\mathrm{u}}}=\mathcal{V}-{\mathcal{V}}_{{\mathrm{l}}} $为剩下的没有标签的节点集. 节点分类的目标是利用图$ G $和$ {\mathcal{V}}_{{\mathrm{l}}} $的标签信息学习可以预测$ {\mathcal{V}}_{{\mathrm{u}}} $标签的映射.
实验操作步骤为:首先,对图数据进行预处理,选择适合的GNN架构学习图中节点的嵌入表示,捕获节点的特征以及它们在图中的结构关系. 之后,引入神经ODE来描述图的动态变化或节点属性的演变过程,并基于学习到的节点表示,使用分类器(如多层感知机)对节点进行分类. 接着,定义适合节点分类任务的损失函数(如交叉熵损失),使用优化算法基于已知的标签信息(即$ {\mathcal{V}}_{{\mathrm{l}}} $)来调整模型参数,并进行交叉验证以确保模型的泛化能力. 随后将训练好的模型应用在$ {\mathcal{V}}_{{\mathrm{u}}} $上对无标签节点进行分类.
节点分类任务中常使用准确率、F1分数等指标评估模型性能.
1)准确率(accuracy)是指正确预测的数量除以总的预测数量. 计算公式为:
$ {a}{c}{c}{u}{r}{a}{c}{y}=\dfrac{\mathrm{正}\mathrm{确}\mathrm{预}\mathrm{测}\mathrm{次}\mathrm{数}}{\mathrm{总}\mathrm{预}\mathrm{测}\mathrm{次}\mathrm{数}}=\dfrac{TP+TN}{TP+TN+FP+FN}. $
在二分类任务中,TP(true positive)与TN(true negative)分别表示模型正确判断为正例和反例的数量,FP(false positive)与FN(false negative)分别表示模型错误判断为正例和反例的数量. 对于多分类问题,可以将单个类别样例视为正例,其他类别样例视为反例,从而构建为多个二分类任务计算公式. 准确率范围在0~1之间,值越高表示模型性能越好.
2)F1分数(F1 score)是精确率$ P $(precision)和召回率$ R $(recall)的调和平均数,计算公式为:
$ F1=2\dfrac{PR}{P+R},\; \; P=\dfrac{TP}{TP+FP},\; \; R=\dfrac{TP}{TP+FN}. $
F1分数范围在0~1之间,值越高表示模型性能越好.
表6总结了图神经ODE模型用于节点分类任务的常用数据集,数据集均为静态图.
1)Cora. 该数据集包含2 708篇机器学习领域的学术论文作为节点,论文的引用关系为边,节点特征用布尔变量表示论文中是否出现某个词语,节点标签是论文所属的次级研究领域,共7种类别.
2)Citeseer. 该数据集包含3 327篇计算机科学领域的学术论文作为节点. 边、特征和标签的含义同Cora,共6种类别.
3)Pubmed. 该数据集包含19 717篇与糖尿病相关的学术论文作为节点,特征为论文中词语的词频-逆文档频率,边和标签的含义同Cora,共3种类别.
4)NELL. 该数据集是一个知识图谱网络,节点为现实世界实体,边为实体间的关系,特征为实体描述文本经词袋模型处理获得,标签为实体类型,共210种类别.
对于给定的动态图$ {\mathcal{G}}_{\mathrm{D}} $或$ {\mathcal{G}}_{\mathrm{C}} $,链接预测的目标是已知某时刻$ t $之前的图信息,对$ t $之后节点间是否存在边进行预测. 具体操作步骤为:首先,对图数据进行预处理,选择适合的GNN学习图中节点的表示. 接着,引入神经ODE来模拟图的动态变化,更新不同时刻的节点表示. 最后,计算节点对之间的相似度或连接概率,进行链接预测. 基于观测到的边或未来时刻边所划分的训练集,定义损失函数并使用优化算法调整模型参数进行交叉验证. 将训练好的模型应用于动态图并推断时刻$ t $之后节点间的边是否存在.
链接预测任务常使用AP(average precision)和AUC(area under curve)等指标评估模型性能.
AP计算在不同召回率水平下精确率的平均值:
$ AP=\displaystyle\sum _{k}({R}_{k}-{R}_{k-1}){P}_{k}, $
AP取值范围为0~1,值越高表示模型性能越好.
AUC计算ROC曲线下与坐标轴围成的面积. ROC曲线描绘了在不同分类阈值下真正类率(即召回率)和假正类率(false positive rate,FPR)之间的关系.
$ FPR=\dfrac{FP}{FP+TN}. $
AUC的取值范围在0~1,值越接近1表示模型性能越好.
表7为用于链接预测任务的常用数据集,均为动态图.
1)Enron. 该数据集中节点为安然公司高级管理层的用户,边为邮件通信.
2)CollegeMsg. 该数据集来自加利福尼亚大学的一个在线社交网络,节点为用户,边为在不同时间戳发送的私人消息互动.
3)Mooc. 该数据集2类节点,分别为Mooc在线平台上的学生和课程,边描述学生与课程之间的互动,如观看视频、提交答案等.
4)Wikipedia. 该数据集节点为用户与Wiki页面,边为用户编辑页面的行为.
5)Reddit. 该数据集节点为Reddit论坛的用户与帖子,边为用户与帖子之间的互动.
6)Taobao. 该数据集节点为电商平台淘宝上的用户与商品,边为用户与商品之间的互动.
对于给定的序列数据$ y=\left\{{y}_{1},… ,{y}_{n}\right\} $,序列预测的目标是通过学习其空间结构和时间动力学,来精确预测序列的未来状态$ \hat{y}=\left\{{\hat{y}}_{1},… ,{\hat{y}}_{n}\right\} $.
具体操作步骤为:首先,收集并预处理序列数据,并构建时间序列的图表示. 选择合适的GNN架构学习时间序列图中节点的时间依赖性嵌入. 其次,引入微分方程模型来描述序列的时间动态,捕捉时间序列的连续变化特性. 再次,将微分方程的动态建模与GNN学习到的时间依赖嵌入结合,用于预测序列的未来状态. 最后,定义适用于序列预测的损失函数,并采用适当的优化算法来调整模型参数.
序列预测任务常使用均方误差、平均绝对误差、平均绝对百分比误差、均方根误差等指标评估模型性能.
1)均方误差(mean square error,MSE)衡量了模型预测值与实际值之间的平均偏差的平方. 其计算公式为:
$ MSE=\dfrac{1}{n}\displaystyle\sum _{i=1}^{n}{({y}_{i}-{\hat{y}}_{i})}^{2}. $
MSE的取值范围在0~+∞,MSE数值越接近0表示模型性能越好.
2)平均绝对误差(mean absolute error,MAE)是预测值与实际值之差的绝对值的平均. 计算公式为:
$ MAE=\dfrac{1}{n}\displaystyle\sum _{i=1}^{n}\left|{y}_{i}-{\hat{y}}_{i}\right|. $
MAE的取值范围在0~+∞,数值越接近0表示模型性能越好.
3)平均绝对百分比误差(mean absolute percentage error,MAPE)是预测值与实际值之间的绝对百分比误差的平均值. 计算公式为:
$ MAPE=\dfrac{100\%}{n}\displaystyle\sum _{i=1}^{n}\left|\dfrac{{y}_{i}-{\hat{y}}_{i}}{{y}_{i}}\right|. $
MAPE的取值范围在0~+∞,数值越接近0表示模型性能越好.
4)均方根误差(root mean square error,RMSE)是预测值与实际值之间差的平方的平均值的平方根. 计算公式为:
$ RMSE=\sqrt{\dfrac{1}{n}\displaystyle\sum _{i=1}^{n}{({y}_{i}-{\hat{y}}_{i})}^{2}}. $
RMSE的取值范围在0~+∞,数值越接近0表示模型性能越好.
表8为图神经ODE模型用于序列预测任务的常用数据集.
PeMS数据集是由加州交通性能测量系统收集的交通数据集,其提供实时的车流量、平均速度和车道占用率等信息. 其中,节点为传感器,边为道路.
通过学习轨迹数据中节点的时空依赖性嵌入,准确预测其未来的位置或轨迹.
具体操作步骤为:首先收集并预处理轨迹的观测数据,构建轨迹的图表示. 之后,选择合适的GNN架构,学习轨迹图中节点的时空依赖性嵌入. 接着,引入微分方程来描述轨迹的动态变化,并结合微分方程的动态建模与GNN学习的时空嵌入,预测轨迹的未来位置. 最后,定义针对轨迹预测任务的损失函数,并使用优化算法调整模型参数.
轨迹预测任务常使用均方误差、平均绝对误差等指标评估模型性能.
这类任务主要采用合成数据集,如由弹簧连接的颗粒、带电粒子、N体带电系统、N体重力系统等来模拟物理系统中粒子的位置、动态行为、轨迹等. 例如通过弹簧相连的粒子系统和带电粒子系统模拟物理系统中粒子的动态行为. 其中,节点和边分别代表系统中的粒子和它们之间的相互作用.
本节将从对图结构的表示能力、对图演化的建模能力、处理不规则观测的能力、计算和存储效率详细地介绍图神经ODE相较于传统GNN的优势.
图神经ODE相较于GNN的一个显著优势是缓解过平滑现象,有助于捕获和表达长程依赖关系. 本节对GNN模型如GCN[65]和GAT[33],以及图神经ODE模型如CGNN[30],GraphCON-GAT[51]和GraphCON-GCN[51]在Cora数据集上以准确率为评价指标进行节点分类任务评估,其中表示维度为64. 结果如图2(a)所示,随着离散层数的增加,GNN模型准确率急剧下降;相应地,随着连续深度的增加,图神经ODE模型的准确率呈现出上升趋势. 图2(b)展示了经不同离散层数/连续深度所得到的狄利克雷能量,该指标越小代表过平滑现象越严重. 随着连续深度的增加,GNN模型的狄利克雷能量显著下降,图神经ODE模型的狄利克雷能量呈现出缓慢上升趋势. 由此可见,图神经ODE能有效缓解过平滑问题,更好地捕获长程相互依赖关系,适用于更深的网络.
图神经ODE能够更加准确地捕获现实系统动力学的演化规律,从而更好地服务于动力学任务. 本节使用GCN作为图结构编码器,并使用传统图循环神经网络来学习有序结构序列之间的时间关系,即得到RNN-GNN,GRU-GNN,LSTM-GNN模型. 设置GNN隐藏维度为5,设置RNN隐藏维度为10,对RNN-GNN,GRU-GNN,LSTM-GNN和NDCN在扩散数据集Grid上以MSE损失为评价指标进行内部插值及外推预测任务评估. 结果如图3所示,随着RNN步数(NDCN模型截止时间)的增加,RNN-GNN,GRU-GNN,LSTM-GNN等传统图循环神经网络的MSE损失始终远高于图神经ODE模型NDCN. 由此可见,图神经ODE模型在推断任意复杂的连续时间潜在动力学上具有卓越的性能.
图神经ODE模型能够摆脱对历史最近快照的强烈依赖,真正捕获网络的时序演化模式. 本节对EvolveGCN,DySAT,STGSN,GRU-GCN和SEGODE在Wiki以及Reddit数据集上以AUC为评价指标进行数据缺失的链接预测任务评估,表示维度为64. 结果如图4所示,在Wiki和Reddit数据集上SEG-ODE链接的AUC值均高于EvolveGCN[37],DySAT[67],STGSN[68],GRU-GNN[69]等时序图神经网络模型. 由此可见,图神经ODE模型对于数据缺失等不规则观测的观测具有优越的性能.
图神经ODE的参数在整个连续深度中共享,极大减少了所需的总参数数量,提高了优化计算与内存使用效率. 本节计算了不同离散层数/连续深度的GNN模型GCN,GAT以及图神经ODE模型CGNN,GraphCON-GAT和GraphCON-GCN在Cora数据集上占用的内存开销,隐藏维度设置为64. 如图5所示,随着离散层数的增加,GNN模型的内存开销线性增长,随着连续深度的增加,图神经ODE模型始终保持低内存开销不变. 由此可见,图神经ODE模型能够显著提升模型的计算和存储效率.
图神经ODE为处理交通网络中的长距离空间-时间依赖问题提供了一种创新解决方案. 交通网络可被视为节点特征不断演化的动态图. 交通流量预测的本质是根据历史观测的动态节点特征和固定的图结构,在图神经ODE的帮助下预测未来的节点特征. GNN能够捕捉交通网络中的空间依赖关系,而利用神经ODE可以动态地整合时间序列变化,二者结合使模型能够根据数据的实际动态变化调整信息聚合过程,更精确地捕捉时间维度的变化趋势和空间维度的相互作用. MTGODE[47],STGODE[46]和ASTGCN[59]等许多工作已开始将图神经ODE用于交通流量预测. 例如,ASTGCN模型同时考虑了局部和全局的时空信息. 该模型通过自适应图卷积和注意力机制获取局部时空信息. 使用TCN块来获得长期时间相关性,并使用ODE建立更深层次的网络以获取全局时空信息,通过对局部和全局时空信息的协力学习,显著提高了交通流量预测的准确性并实现精准的交通流量预测,能够为智能交通系统的决策支持提供科学依据,有助于城市交通管理、路线规划和交通拥堵预防等多方面的应用.
图神经ODE为模拟和预测复杂物理系统中的动态行为开辟了新领域. 通过将GNN的强大表示能力与神经ODE的连续时间动态模型相结合,不仅可以精确捕捉系统中多个对象之间的复杂相互作用,还能在连续时间内模拟这些相互作用如何随时间演变. 例如,Liu等人[53]提出的PINGO模型,首先采用等变GNN作为特征提取器处理物理系统中对象的相互作用,进而利用神经ODE函数处理GNN获取的信息,模拟系统中对象状态随时间的连续演化. 这种方法的灵活性和泛化能力使其能适用于从模拟分子动力学到模拟宇宙动力学等广泛的应用,提供了对物理系统演化过程深入理解的强大工具. 特别是,该方法的理论框架为解的唯一性和稳定性提供了保障,这大大增加了模型的可靠性. 因此,图神经ODE为科学研究和工程应用中的进一步探索和创新奠定了新的方法基础.
图神经ODE在处理复杂气候系统方面具有独特优势. 可以通过利用GNN捕捉地理位置间的空间依赖性,并借助神经ODE在连续时间上建模气候系统状态的动态变化. 例如,Hwang等人[58]建立神经ODE模型在气候建模中用于模拟扩散过程和固有不确定性. 该模型用一个编码层处理节点特征矩阵,并构建热容生成层生成热容系数以模拟环境变量热容特性. 接着用神经网络学习不确定性,并采用了离散的拉普拉斯算子建模网格网络上的扩散过程,进而建立了神经ODE函数用于预测气候因子. 这种方法整合了时间和空间维度的复杂关系,不仅提高了预测的准确性,也增强了模型的灵活性和解释性. 因此,图神经ODE为气候科学家提供了一种高效的工具,以更准确地预测温度、降雨量等气候变量,及时预警极端气候事件,为气候变化研究和政策制定提供支持.
图神经ODE在疾病、舆论等的传播预测研究中展示出巨大的应用潜力. 可以通过GNN捕捉地理或社交网络中的空间依赖性,并利用ODE模拟疾病、舆论在这些网络中的传播动力学. 图神经ODE能够有效地模拟和预测诸如COVID-19的流行病在复杂人口网络中的传播情况[48]. 这种模型不仅可以帮助研究人员准确预测疾病、舆论等的未来传播趋势,还可以评估对病原、谣言等的管控效果,为制定公共政策和资源分配提供科学依据. 尽管传播预测对数据质量要求严格,但随着技术的不断进步,图神经ODE在公共卫生、网络安全等领域的应用前景仍然十分广阔.
图神经ODE在推荐系统中的应用提供了一种创新的方法,能够动态地捕捉和理解用户偏好及其随时间的演变. 可以通过GNN的能力,捕获用户和商品之间的复杂偏好关系,并结合神经ODE的能力,模拟这些偏好随时间的连续动态变化. 图神经ODE能够有效地处理推荐系统中的关键挑战,包括对用户行为的不规则间隔进行自然处理,以及捕捉用户-商品交互图中的高阶协同信号. 例如,Choi等人[49]引入图神经ODE来改善传统协同过滤和图卷积网络方法. 该方法分别定义了代表用户和商品表示随时间变化的ODE函数,用于捕捉用户和商品之间的相互作用. 然后,引入了一组可学习的时间点,用于提取关键状态,在用户和商品表示的演化过程中整合这些状态,形成最终的用户和商品表示. 该方法形成了用户和产品相互依赖和共同演化的双重ODE模型,能精准捕捉用户偏好和商品特性的动态变化,提高了推荐的准确性,增强了模型对时间信息的灵活性. 因此,图神经ODE在理解和预测动态变化的用户偏好方面展现出独特的优势,有潜力成为推荐系统的强大工具.
图神经ODE在因果推断[70]领域的应用提供了一种强大的框架,特别适用于分析和理解多智能体动态系统中的复杂相互作用和依赖性. 图神经ODE能够在连续时间内精准地建模各单元之间的相互影响及其随时间的演变. 图神经ODE通过细致地模拟单元间的动态相互作用,能够估计在不同干预措施下的连续时间反事实结果,从而为评估干预策略提供了强有力的支持. 例如,Jiang等人[50]建立了基于图神经ODE在多智能体动态系统中估计连续时间反事实结果的因果模型. 该模型从初始观测数据中学习得到每个节点的初始潜在表示,随后以节点接受的干预作为额外的输入,通过ODE框架动态建模节点状态的演化,从而利用干预诱导图神经ODE学习节点的连续潜在表示轨迹. 这一方法解决了传统方法难以克服的多源混杂因素、干扰不平衡及系统连续性等问题,有望推动医疗干预、环境政策、社会网络等领域中准确因果关系理解的前进,在现代因果推断研究中具有重要价值和潜力.
高阶图结构和空间导数的整合可以引入新的模型架构,以更好地捕捉图数据的复杂空间关系,从而大幅度提升模型的能力,但对模型复杂性和计算开销提出了巨大挑战. 如何有效地表征和计算高阶邻接信息,以及如何准确地将空间导数融入图神经ODE模型中,是需要解决的关键问题. 为了应对这些挑战,需要开发高效的算法和优化技术,以降低计算负担并提高模型的学习能力.
现有的图神经ODE模型主要聚焦于对静态图和动态图连续深度建模. 在更复杂的图数据,如多模态图、超图等之上,如何有效地实现连续深度建模仍是未被探究的问题. 因此,未来的研究可以开发新的数学框架和算法,以支持对各种复杂图结构的连续建模,探索新的图表示方法,提升模型应对多模态数据和高阶关系中复杂特性的能力.
图神经ODE通过在连续的深度维度上建模图数据,允许模型以无限的层次来逼近任何可微分的动态系统,能够更细致地捕捉到图结构数据随时间的演化. 但是,图神经ODE的表达能力仍有巨大的提升空间. 图神经ODE不仅面临数值稳定性的挑战,其表达能力也受到对图结构学习能力的限制. 传统的GNN通过聚合邻居节点的信息来更新节点状态,而图神经ODE需要在整个连续深度维度上学习这种更新规则. 这要求模型不仅能够捕获节点间的即时相互作用,还能够理解这些相互作用如何随时间演变. 因此,设计能够有效捕获图数据中时间依赖性和复杂性的图神经ODE架构是提高模型表达能力的关键. 此外,未来的研究还可以探索通过多尺度建模、引入更高级的数值求解策略,以及结合强化学习等其他机器学习技术来进一步增强图神经ODE的表达能力.
图结构数据易受到对抗性攻击的影响,即轻微修改图的结构或特征可能会导致模型性能显著下降. 处理动态变化的图数据时,保障模型的稳定性和可靠性显得更为重要. 因此,图神经ODE模型在维持对动态数据建模灵活性的同时,也需要增强对这些攻击的抵抗能力. 而大规模图数据不仅对计算资源的需求高,也对模型的鲁棒性提出了更大的挑战. 设计既能高效处理大规模图数据,又能确保高鲁棒性的图神经ODE模型,是一个开放且富有挑战的问题. 在这方面,研究人员可以探索设计更加稳健的图神经ODE模型,通过引入对抗性训练、正则化技术等手段来提高模型的鲁棒性. 此外,对于图结构数据的特性和攻击方式进行深入分析,也有助于指导鲁棒性模型的设计和改进.
图神经ODE模型的内部机制变得越来越复杂,这虽然给模型带来能力的提升,但也直接影响了模型的可解释性. 图神经ODE模型的状态变化反映了图上的节点和边随时间的演化,理解和解释这些变化背后的因果机制是一个复杂的任务. 尤其是在多步长的连续时间演化中,识别关键驱动因素变得更加困难. 为了解决上述挑战,需要开发专门用于解释图神经ODE模型的新工具和新技术,包括基于模型内部机制的白盒解释方法与依赖于模型输出的黑盒解释方法等. 此外,由于实际应用领域的多样性及相应建模的复杂性,图神经ODE模型的可解释性研究需要综合数学、计算机科学、物理学以及特定应用领域的各种知识,以更好地理解模型行为.
近年来,海量参数的大语言模型在各类自然语言任务中展现出媲美甚至超出人类的水平. 然而在图领域,传统GNN架构受限于过平滑等问题难以深度堆叠,导致具备高精度、高泛化能力的图大模型仍未出现. 图神经ODE为研制图大模型提供了新思路. 一方面,图神经ODE本身具备连续深度,其高效的参数利用率也为深度堆叠模型提供了便利,这是图大模型高精度的基础;另一方面,图神经ODE能够自适应地学习图数据中的动力学,具备良好的推理能力,能够灵活的泛化于各种图领域任务.
近年来,将神经ODE应用于图相关任务成为了研究热点. 本文深入综述了图神经ODE模型:首先回顾了引入图神经ODE的背景,并介绍了相关的核心概念. 其次创新性地提出了一种新的分类方法,根据图的不同特性,将研究工作分为静态图ODE机制扩展、离散动态图神经ODE、连续动态图神经ODE三大类. 对于每一类方法,本文总结了一个通用框架并对ODE与GNN结合的作用机制进行了系统性的讨论. 接着,详细介绍了该领域常见的验证任务和相应的数据集,并提供了相关论文和开源代码链接. 最后,探讨了图神经ODE的应用现状、挑战以及未来的发展趋势,旨在为未来的研究提供宝贵的参考和启示.
作者贡献声明:焦鹏飞提出了论文框架、写作思路,并参与论文修订;陈舒欣撰写了论文的核心部分;郭翾修订论文;何东晓和刘栋提出指导意见并修改论文.

参考文献 引证文献
排序方式:
1
徐冰冰,岑科廷,黄俊杰,等. 图卷积神经网络综述[J]. 计算机学报,2020,43(5):755−780
Xu Bingbing, Cen Keting, Huang Junjie, et al. A survey on graph convolutional network[J]. Chinese Journal of Computers, 2020, 43(5): 755−780 (in Chinese)
2
马帅,刘建伟,左信. 图神经网络综述[J]. 计算机研究与发展,2022,59(1):47−80
Ma Shuai, Liu Jianwei, Zuo Xin. Survey on graph neural network[J]. Journal of Computer Research and Development, 2022, 59(1): 47−80 (in Chinese)
3
Wu Zonghan, Pan Shirui, Chen Fengwen, et al. A comprehensive survey on graph neural networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2020, 32(1): 4−24
4
Liu Juncheng, Hooi B, Kawaguchi K, et al. Scalable and effective implicit graph neural networks on large graphs[C/OL]//Proc of the 12th Int Conf on Learning Representations. 2024[2024-05-16].https://openreview.net/forum?id=QcMdPYBwTu
5
Xu Keyulu, Hu Weihua, Leskovec J, et al. How powerful are graph neural networks?[C/OL]//Proc of the 7th Int Conf on Learning Representations. 2019[2024-05-16].https://openreview.net/forum?id=ryGs6iA5Km
6
Gasteiger J, Bojchevski A, Günnemann S. Predict then propagate: Graph neural networks meet personalized pagerank[C/OL]//Proc of the 7th Int Conf on Learning Representations. 2019[2024-05-16].https://openreview.net/forum?id=H1gL-2A9Ym
7
Hajiramezanali E, Hasanzadeh A, Narayanan K, et al. Variational graph recurrent neural networks[C]//Proc of the 33rd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2019, 32: 10701−10711
8
Xu Da, Ruan Chuanwei, Korpeoglu E, et al. Inductive representation learning on temporal graphs[C/OL]//Proc of the 8th Int Conf on Learning Representations. 2020[2024-05-16].https://openreview.net/forum?id=rJeW1yHYwH
9
Wu Yongji, Lian Defu, Xu Yiheng, et al. Graph convolutional networks with Markov random field reasoning for social spammer detection[C]//Proc of the 34th AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2020, 34(1): 1054−1061
10
Zhang Si, Tong Hanghang, Xu Jiejun, et al. Graph convolutional networks: A comprehensive review[J]. Computational Social Networks, 2019, 6(1): 1−23
11
Ying Rex, He Ruining, Chen Kaifeng, et al. Graph convolutional neural networks for web-scale recommender systems[C]//Proc of the 24th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2018: 974−983
12
王磊,熊于宁,李云鹏,等. 一种基于增强图卷积神经网络的协同推荐模型[J]. 计算机研究与发展,2021,58(9):1987−1996
Wang Lei, Xiong Yuning, Li Yunpeng, el al. A collaborative recommendation model based on enhanced graph convolutional neural network[J]. Journal of Computer Research and Development, 2021, 58(9): 1987−1996 (in Chinese)
13
Wang Shoujin, Hu Liang, Wang Yan, et al. Graph learning based recommender systems: A review[C]//Proc of the 30th Int Joint Conf on Artificial Intelligence. San Francisco, CA: IJCAI, 2021: 4644−4652
14
Fout A, Byrd J, Shariat B, et al. Protein interface prediction using graph convolutional networks[C]//Proc of the 31st Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2017, 30: 6533−6542
15
Borgwardt K M, Ong C S, Schönauer S, et al. Protein function prediction via graph kernels[J]. Bioinformatics, 2005, 21(suppl_1): i47−i56
16
Guo Shengnan, Lin Youfang, Feng Ning, et al. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting[C]//Proc of the 33rd AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2019: 922−929
17
Yu Bing, Yin Haoteng, Zhu Zhanxing. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting[C]//Proc of the 28th Int Joint Conf on Artificial Intelligence. San Francisco, CA: IJCAI, 2018: 3634−3640
18
杜圣东,李天瑞,杨燕,等. 一种基于序列到序列时空注意力学习的交通流预测模型[J]. 计算机研究与发展,2020,57(8):1715−1728
Du Shengdong, Li Tianrui, Yang Yan, et al. A sequence-to-sequence spatial-temporal attention learning model for urban traffic flow prediction[J]. Journal of Computer Research and Development, 2020, 57(8): 1715−1728 (in Chinese)
19
Bodnar C, Di Giovanni F, Chamberlain B, et al. Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in GNNS[C]//Proc of the 36th Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2022, 35: 18527−18541
20
Rusch T K, Bronstein M M, Mishra S. A survey on oversmoothing in graph neural networks[J]. arXiv preprint, arXiv: 2303.10993, 2023
21
Rong Yu, Huang Wenbing, Xu Tingyang, et al. DropEdge: Towards deep graph convolutional networks on node classification[C/OL]//Proc of the 8th Int Conf on Learning Representations. 2020 [2024-05-16].https://openreview.net/forum?id=Hkx1qkrKPr
22
Zhao L, Akoglu L. PairNorm: Tackling oversmoothing in GNN[C]//Proc of the 8th Int Conf on Learning Representations. 2020[2024-05-16].https://openreview.net/forum?id=rkecl1rtwB
23
Xu Keyulu, Li Chengtao, Tian Yonglong, et al. Representation learning on graphs with jumping knowledge networks[C]//Proc of the 35th Int Conf on Machine Learning. New York: PMLR, 2018: 5453−5462
24
Barzel B, Liu Yangyu, Barabási A L. Constructing minimal models for complex system dynamics[J]. Nature Communication, 2015, 6(1): 7186
25
Alvarez-Rodriguez U, Battiston F, de Arruda G F, et al. Evolutionary dynamics of higher-order interactions in social networks[J]. Nature Human Behaviour, 2021, 5(5): 586−595
26
Zhang Yanfu, Gao Shangqian, Pei Jian, et al. Improving social network embedding via new second-order continuous graph neural networks[C]//Proc of the 28th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2022: 2515−2523
27
Hou Jinlin, Guo Xuan, Liu Jiye, et al. Structure-enhanced graph neural ODE network for temporal link prediction[C]//Proc of the 33rd Int Conf on Artificial Neural Networks. Berlin: Springer, 2023: 563−575
28
Fey M, Lenssen J E, Weichert F, et al. Gnnautoscale: Scalable and expressive graph neural networks via historical embeddings[C]//Proc of the 38th Int Confon Machine Learning. New York: PMLR, 2021: 3294−3304
29
Chen R T Q, Rubanova Y, Bettencourt J, et al. Neural ordinary differential equations[C]//Proc of the 32nd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2018, 31: 6572−6583
30
Xhonneux L P, Qu Meng, Tang Jian. Continuous graph neural networks[C]//Proc of the 37th Int Conf on Machine Learning. New York: PMLR, 2020: 10432−10441
31
Zang Chengxi, Wang Fei. Neural dynamics on complex networks[C]//Proc of the 26th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2020: 892−902
32
Wang Zihui, Yang Peizhen, Fan Xiaoliang, et al. ConTIG: Continuous representation learning on temporal interaction graphs[J]. Neural Networks, 2024, 172: 106151
33
Veličković P, Cucurull G, Casanova A, et al. Graph attention networks[C/OL]//Proc of the 6th Int Conf on Learning Representations. 2018[2024-05-16].https://openreview.net/forum?id=rJXMpikCZ
34
Rossi E, Chamberlain B, Frasca F, et al. Temporal graph networks for deep learning on dynamic graphs[C]//Proc of the ICML 2020 Workshop on Graph Representation Learning and Beyond. New York: PMLR, 2020: 1−9
35
Ma Yao, Guo Ziyi, Ren Zhaocun, et al. Streaming graph neural networks[C]//Proc of the 43rd Int ACM SIGIR Conf on Research and Development in Information Retrieval. New York: ACM, 2020: 719−728
36
Zhou Hongkuan, Zheng Da, Nisa I, et al. TGL: A general framework for temporal gnn training on billion-scale graphs[C]//Proc of the 48th Int Conf on Very Large Databases. New York: VLDB Endowment, 2022, 15(8): 1572−1580
37
Pareja A, Domeniconi G, Chen Jie, et al. EvolveGCN: Evolving graph convolutional networks for dynamic graphs[C]//Proc of the 34th AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2020: 5363−5370
38
Taheri A, Gimpel K, Berger-Wolf T. Learning to represent the evolution of dynamic graphs with recurrent models[C]//Proc of Companion Proc of the 2019 World Wide Web Conf. New York: ACM, 2019: 301−307
39
Wang Yanbang, Li Pan, Bai Chongyang, et al. Generic representation learning for dynamic social interaction[C]//Proc of the 26th ACM SIGKDD Conf on Knowledge Discovery and Data Mining Workshop. New York: ACM, 2020: 1−9
40
You Jiaxuan, Du Tianyu, Leskovec J. ROLAND: Graph learning framework for dynamic graphs[C]//Proc of the 28th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2022: 2358−2366
41
Pontryagin L S. Mathematical Theory of Optimal Processes[M]. London: Routledge, 2018
42
Rubanova Y, Chen R T Q, Duvenaud D K. Latent ordinary differential equations for irregularly-sampled time series[C]//Proc of the 33rd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2019, 32: 5320−5330
43
Dupont E, Doucet A, Teh Y W. Augmented neural ODEs[C]//Proc of the 33rd Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2019, 32: 3140−3150
44
Biswas B N, Chatterjee S, Mukherjee S P, et al. A discussion on Euler method: A review[J]. Electronic Journal of Mathematical Analysis and Applications, 2013, 1(2): 2090−2792
45
Dormand J R, Prince P J. A family of embedded Runge-Kutta formulae[J]. Journal of Computational and Applied Mathematics, 1980, 6(1): 19−26
46
Fang Zheng, Long Qingqing, Song Guojie, et al. Spatial-temporal graph ode networks for traffic flow forecasting[C]//Proc of the 27th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2021: 364−373
47
Jin Ming, Zheng Yu, Li Yuanfang, et al. Multivariate time series forecasting with dynamic graph neural ODES[J]. IEEE Transactions on Knowledge and Data Engineering, 2023, 35(9): 9168−80
48
Huang Zijie, Sun Yizhou, Wang Wei. Coupled graph ODE for learning interacting system dynamics[C]//Proc of the 27th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2021: 705−715
49
Choi J, Jeon J, Park N. LT-OCF: Learnable-time ode-based collaborative filtering[C]//Proc of the 30th ACM Int Conf on Information and Knowledge Management. New York, ACM, 2021: 251−260
50
Jiang Song, Huang Zijie, Luo Xiao, et al. CF-GODE: Continuous-time causal inference for multi-agent dynamical systems[C]//Proc of the 29th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2023: 997−1009
51
Rusch T K, Chamberlain B, Rowbottom J, et al. Graph-coupled oscillator networks[C]//Proc of the 39th Int Conf on Machine Learning. New York: PMLR, 2022: 18888−18909
52
Huang Zijie, Zhao Wanjie, Gao Jingdong, et al. TANGO: Time-reversal latent GraphODE for multi-agent dynamical systems[C/OL]//Proc of NeurIPS 2023 Workshop on the Symbiosis of Deep Learning and Differential Equations. 2023[2024-05-16]. https://openreview.net/forum?id=8AYetbupoZ
53
Liu Yang, Cheng Jiashun, Zhao Haihong, et al. Physics-inspired neural graph ODE for long-term dynamical simulation[J]. arXiv preprint, arXiv: 2308.13212, 2023
54
Schober M, Särkkä S, Hennig P. A probabilistic model for the numerical solution of initial value problems[J]. Statistics and Computing, 2019, 29(1): 99−122
55
Poli M, Massaroli S, Park J, et al. Graph neural ordinary differential equations [J]. arXiv preprint, arXiv: 1911.07532, 2019
56
Huang Zijie, Sun Yizhou, Wang Wei. Learning continuous system dynamics from irregularly-sampled partial observations[C]//Proc of the 34th Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2020, 33: 16177−16187
57
Huang Zijie, Sun Yizhou, Wang Wei. Generalizing graph ODE for learning complex system dynamics across environments[C]//Proc of the 29th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2023: 798−809
58
Hwang J, Choi J, Choi H, et al. Climate modeling with neural diffusion equations[C]//Proc of 2021 IEEE Int Conf on Data Mining. Piscataway, NJ: IEEE, 2021: 230−239
59
Chen Yibi, Qin Yunchuan, Li Kenli, et al. Adaptive spatial-temporal graph convolution networks for collaborative local-global learning in traffic prediction[J]. IEEE Transactions on Vehicular Technology, 2023, 72(10): 12653−12663
60
Guo Jiayan, Zhang Peiyan, Li Chaozhuo, et al. Evolutionary preference learning via graph nested GRU ODE for session-based recommendation[C]//Proc of the 31st ACM Int Conf on Information and Knowledge Management. New York: ACM, 2022: 624−634
61
Luo Xiao, Yuan Jingyang, Huang Zijie, et al. HOPE: High-order graph ODE for modeling interacting dynamics[C]//Proc of the 40th Int Conf on Machine Learning. New York: PMLR, 2023: 23124−23139
62
Jin Ming, Li Yuanfang, Pan Shirui. Neural temporal walks: Motif-aware representation learning on continuous-time dynamic graphs[C]//Proc of the 36th Int Conf on Neural Information Processing Systems. New York: Curran Associates, Inc, 2022, 35: 19874−19886
63
Qin Yifang, Ju Wei, Wu Hongjun, et al. Learning graph node for continuous-time sequential recommendation [J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(7): 1−14
64
Gao Zhihan, Wang Hao, Wang Yuyang, et al. Probabilistic continuous-time whole-graph forecasting[C]//Proc of the 8th SIGKDD Int Workshop on Mining and Learning from Time Series. New York: ACM, 2022: 1−14
65
Kipf T N, Welling M. Semi-supervised classification with graph convolutional networks[C/OL]//Proc of the 5th Int Conf on Learning Representations. 2017[2024-05-16]. https://openreview.net/forum?id=SJU4ayYgl
66
Skarding J, Gabrys B, Musial K. Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey[J]. IEEE Access, 2021, 9: 79143−79168
67
Sankar A, Wu Yanhong, Gou Liang, et al. DySAT: Deep neural representation learning on dynamic graphs via self-attention networks[C]//Proc of the 13th Int Conf on Web Search and Data Mining. New York: ACM, 2020: 519−527
68
Min Shengjie, Gao Zhan, Peng Jing, et al. STGSN—a spatial temporal graph neural network framework for time-evolving social networks[J]. Knowledge-Based Systems, 2021, 214: 106746
69
Gao Jianfei, Ribeiro B. On the equivalence between temporal and static equivariant graph representations[C]//Proc of the 39th Int Conf on Machine Learning. New York: PMLR, 2022: 7052−7076
70
Yao Liuyi, Chu Zhixuan, Li Sheng, et al. A survey on causal inference[J]. ACM Transactions on Knowledge Discovery from Data, 2021, 15(5): 1−46
2024年第61卷第8期
95
45
0
引用本文
BibTeX
文章信息
doi: 10.7544/issn1000-1239.202440192
  • 接收时间:2024-03-16
  • 首发时间:2026-11-17
  • 出版时间:2024-08-01
补充材料
相关文章
文章信息
作者
出版历史
  • 收稿日期:2024-03-16
  • 修回日期:2024-05-15
基金
作者信息
    1杭州电子科技大学网络空间安全学院 杭州 310018
    2杭州电子科技大学卓越学院 杭州 310018
    3天津大学智能与计算学部 天津 300350
    4河南师范大学计算机与信息工程学院 河南新乡 453007

通讯作者:

何东晓()
参考文献
分享链接
https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202440192
分享至
全文二维码

扫描看全文

引用本文
BibTeX
本文的引用情况
2种不同金属材料的力学参数
科
Family
属数
Number of
genus
种数
Number of
species
占总种数比例
Percentage of
total species (%)
属
Genus
种数
Number of
species
占总种数比例
Percentage of total
species (%)
鹅膏菌科Amanitaceae 2 11 5.26 鹅膏菌属 Amanita 10 4.78
小菇科 Mycenaceae 2 12 5.74 丝盖伞属 Inocybe 5 2.39
多孔菌科 Polyporaceae 8 14 6.70 蜡蘑属 Laccaria 5 2.39
红菇科 Russulaceae 3 23 11.00 小皮伞属 Marasmius 6 2.87
小菇属 Mycena 11 5.26
光柄菇属 Pluteus 5 2.39
红菇属 Russula 17 8.13
栓菌属 Trametes 5 2.39
关闭全屏