图神经网络浅析

Overview of Graph Neural Network

近年来,深度学习已经彻底改变许多机器学习任务,从图像分类和视频处理到语音识别和自然语言理解。这些任务中的数据通常用欧几里得空间表示。然而,越来越多的应用需要使用非欧几里得域的数据,并将其表示为包含对象之间关系的图。例如,对物理系统进行建模、学习分子指纹和预测蛋白质界面,这些任务需要从输入的图中学习模型。一些其他领域,如文本、图像等非结构化数据的学习,以及句子依赖树、图像场景图等提取结构的推理,也需要图的推理模型。近年来,关于图的深度学习方法的研究层出不穷。图神经网络(GNN)是一种连接主义模型,它通过图节点之间的消息传递来捕获图的依赖性。与标准的神经网络不同,图神经网络保持一种状态,它可以以任意深度表示来自其邻域的信息。对图神经网络模型的发展和基本原理进行概述,对其应用进行系统的分类,并提出四个有待进一步研究的问题。

In recent years,deep learning has radically changed many machine learning tasks,from image classification and video processing to speech recognition and natural language understanding.Data in these tasks are usually represented in Euclidean space.However,more and more applications need to use data from non-Euclidean space and represent them as graphs containing relationships between objects.For exam?ple,modeling physical systems,learning molecular fingerprints,and predicting protein interfaces require learning models from input graphs.Some other fields,such as the learning of non-structural data such as text and image,and the reasoning of extracting structures such as sentence dependency tree and image scene graph,also need the reasoning model of graph.In recent years,there have been many studies on the deep learning methods of graphs.Graph Neural Network(GNN)is a connectionist model that captures graph dependencies by message passing between graph nodes.Unlike standard neural networks,graph neural networks maintain a state in which information from their neighborhoods can be represented at any depth.Summarizes the development and basic principle of graph neural network model,classifies its application systematically,and puts forward four problems to be further studied.