图神经网络前沿进展与应用

图结构数据是现实生活中广泛存在的一类数据形式.宏观上的互联网、知识图谱、社交网络数据,微观上的蛋白质、化合物分子等都可以用图结构来建模和表示.由于图结构数据的复杂性和异质性,对图结构数据的分析和处理一直是研究界的难点和重点.图神经网络(Graph Neural Network,GNN)是近年来出现的一种利用深度学习直接对图结构数据进行学习的框架,其优异的性能引起了学者高度的关注和深入的探索.通过在图中的节点和边上制定一定的策略,GNN将图结构数据转化为规范而标准的表示,并输入到多种不同的神经网络中进行训练,在节点分类、边信息传播和图聚类等任务上取得优良的效果.与其他图学习算法相比较,GNN能够学习到图结构数据中的节点以及边的内在规律和更加深层次的语义特征.由于具有对图结构数据强大的非线性拟合能力,因此在不同领域的图相关问题上,GNN都表现出更高的准确率和更好的鲁棒性.本文在现有GNN研究的基础上,首先概述了GNN的出现历程,并介绍了相关概念和定义.之后本文着重讨论和对比了GNN中的各种算法框架,包括核心思想、任务划分、学习方式、优缺点、适用范围、实现成本等.此外,本文对GNN算法在多个不同领域下的应用场景进行了详细的阐述,将GNN与其他图学习算法的优缺点作了联系和比较.针对存在的一些问题和挑战,本文勾画了GNN的未来方向和发展趋势,最后对全文进行了全面而细致的总结.

Advances and Applications in Graph Neural Network

As is known to all,Graph-structure data is a kind of data form widely existing in real life.Internet network,knowledge graph,social network data in macro perspective,together with protein,compound molecules data in micro perspective,are all can be modeled and represented by graph-structure.Because graph-structure data has complexity and heterogeneity attribute,the analysis and processing of graph-structure data have always been a difficulty in research community.The researchers have been studying the property information and topological structure information in graph and try to find out a way or method to learn and explore the graph automatically.In order to solve the problems above,Graph Neural Network(GNN)appears as a kind of framework that uses deep learning to learn graph-structure data directly in recent years.On the one hand,the excellent performance of GNN has aroused high attention and deep exploration in research community.GNN transforms the graph-structure data into standard representation by making a series of certain strategies on the multifarious nodes and edges of the graph,and then the representation can be input into a variety of different artificial neural networks for training,and achieves excellent results in tasks such as node classification,edge information dissemination,graph clustering and so on.On the other hand,when it is compared with other graph learning algorithms,GNN can learn the internal rules and semantic features of node and edge features in graph-structure data.Because it has a strong nonlinear fitting ability to graph-structure data,GNN has higher accuracy and better robustness on graph-structure related problems in different fields.To make it more suitable and efficient for specific applications,there is a great deal of variants of GNN algorithm and framework are proposed in past few years.Based on the existing GNN research,this paper first summarizes the history of GNN,and introduces the related concepts and definitions.After an overview of GNN theory,we then focus on the discussion and comparison of various algorithms in GNN,including the core idea,task division,types of graphs,activation function,different dataset,advantages and disadvantages,the scope of application,implementation costs,learning methods and benchmark network.We give a novel classification and divide GNN into five different artificial neural networks.In addition,the application of GNN algorithm in many different fields is described in details such as natural language processing,molecule graph generation and so on.This paper gives an introduction to the other kinds of graph learning algorithm which are recognized as network embedding and graph kernel.We compare the advantages and disadvantages between GNN and network embedding as well as graph kernel.Although GNN has been very popular over past years,these two kinds of graph learning algorithm are also to be proved competitive in some tasks.In view of the existing problems and challenges,this paper outlines the future direction and development trend of GNN,which includes depth of artificial neural network,dynamics,receptive field of GNN,the fusion of multi artificial neural networks,and the combination between artificial network embedding and GNN.Last but not least,we make a comprehensive and detailed summary of the full text.