图神经网络技术研究综述  

Research review of graph neural network technology

作　　者：李甜甜 张荣梅[1] 张佳惠 LI Tian-tian;ZHANG Rong-mei;ZHANG Jia-hui(College of Information Technology, Hebei University of Economics and Business, Shijiazhuang Hebei 050061, China)

机构地区：[1]河北经贸大学信息技术学院,河北石家庄050061

出　　处：《河北省科学院学报》2022年第2期1-13,共13页Journal of The Hebei Academy of Sciences

摘　　要：近年来,深度学习已成功应用于图像处理、自然语言理解等领域,在图像、语音、文本等形式的数据上获得较好效果。但深度学习一直无法很好地对于图形式的非结构化数据进行有效的适配。而作为一类主要用于描述关系的通用数据表示方法,图数据在产业界有着更加广阔的应用场景,例如社交网络、电子购物、物联网、生物制药等场景中数据多以图的形式出现。于是,将深度学习技术迁移到图数据处理的图神经网络技术于2005年被提出,并且受到了非常广泛的关注。本文对图神经网络技术进行综述,首先梳理了图神经网络相关的背景知识,介绍了四个主要的图神经网络模型原理、主要应用领域以及开放资源,最后分析了图神经网络面临的主要问题。Recently,deep learning has been successfully applied in image processing,natural language understanding and other fields,and has achieved great improvement on data in the form of images,speech and text.However,deep learning has not been able to adapt well to unstructured data in the form of graph effectively.As a general data representation method that mainly describes relationships,graph data has broader application scenarios in industry,such as social networks,e-shopping,Internet of Things,biopharmaceuticals and other scenarios where data are mostly in the form of graphs.Therefore,the graph neural network technology,which transfers deep learning technology to graph data processing,was proposed in 2005 and has gained wide attention.In this paper,the technology of graph neural network is summarized.Firstly,the background knowledge of graph neural network are described.Then,the four main graph neural network model principles,main application domains and open resources are introduced.Finally the main problems faced by graph neural network are analyzed.

关 键 词：图神经网络 GCN GAT GraphSAGE GGNN 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]