图神经架构搜索综述  

Graph Neural Architecture Search:A Survey

作　　者：张子威 王鑫 朱文武 ZHANG Zi-Wei;WANG Xin;ZHU Wen-Wu(Department of Computer Science and Technology,Tsinghua University,Beijing 100084)

机构地区：[1]清华大学计算机科学与技术系,北京100084

出　　处：《计算机学报》2023年第7期1532-1552,共21页Chinese Journal of Computers

基　　金：科技创新2030——“新一代人工智能重大项目”(2020AAA0106300);国家自然科学基金项目(62250008,62222209,62102222,61936011,62206149);北京信息科学与技术国家研究中心(BNR2023RC01003);网络多媒体北京市重点实验室、博士后创新人才支持计划(BX20220185);中国博士后科学基金面上项目(2022M711813)资助。

摘　　要：图数据可以广泛建模事物之间的复杂关系.小到蛋白质中的分子与氨基酸结构,大到世界范围的物流与交通网络;从人类社会的社交网络,到信息空间的互联网,均可统一表示为图数据的形式.图数据中蕴藏着巨大的研究与应用价值.图神经网络是过去几年中图数据上进行机器学习的主要范式.通过在图数据的链接关系上重新定义神经网络架构并实现端到端的学习,图神经网络可以有效处理节点分类、链接预测、图分类等多种图数据分析与挖掘任务.然而,由于图数据的复杂性、图任务的多样性以及图神经网络的复杂程度,人工设计最优的图神经网络架构变得越来越困难,且无法适应开放变化环境.图神经架构搜索,旨在自动化设计针对特定数据集与任务的最优图神经网络架构,应运而生并逐渐受到了学术界和工业界的关注.在本文中,我们对图神经架构搜索这一快速发展的新兴领域进行综述.特别地,我们系统总结并梳理了目前已公开发表的四十余篇图神经架构搜索算法,并从搜索空间、搜索策略、模型性能评估策略以及其他特点对已有算法进行了全面的分类、对比与评述,并从实验上对上述算法进行了归纳.此外,我们还对近期的图神经架构搜索研究趋势进行了评述.最后,我们分享了对图神经架构搜索未来研究方向的看法.Graph data can generally model the complex relationships between entities.From as small as the molecular and amino acid structures in proteins,to as large as the worldwide logistics and transportation networks;from the social networks in the human society to the Internet in the information space,all these data can be uniformly represented as graphs.Huge research and application values exist underlying the graph data.Graph neural networks are the dominant paradigm for machine learning on graphs over the past few years.By redefining neural network architectures on the link relationship of graph data and realizing end-to-end learning paradigms,graph neural networks can effectively handle a variety of graph analytical and mining tasks such as node classification,link prediction,and graph classification.However,due to the complexity of graph data,the diversity of graph tasks,and the complex architecture of graph neural networks,it becomes increasingly difficult to manually design the optimal graph neural network architectures,failing to adapt to open and changing environments.Graph neural architecture search,which aims to automate the design of optimal graph neural network architectures for specific data sets and tasks,comes into being and has attracted considerable attention from both academia and industry.In this paper,we provide a comprehensive review for the rapidly evolving and emerging field of graph neural architecture search.In particular,we systematically review and summarize more than 40 published graph neural architecture search algorithms,and comprehensively classify,compare and comment on the existing algorithms from the search space,the search strategy,the model evaluation strategy,and other characteristics of the models.We also summarize the above algorithms from the experimental point of view.In addition,we analyze recent trends in graph neural architecture search studies.Finally,we share our insights on the future research direction of graph neural architecture search.

关 键 词：图神经网络 神经架构搜索 图机器学习 自动机器学习 人工智能 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]