| 摘要: |
| 图神经网络(GNN)是一类基于深度学习的处理图域信息的方法, 它通过将图广播操作和深度学习算法结合, 可以让图的结构信息和顶点属性信息都参与到学习中, 在顶点分类、图分类、链接预测等应用中表现出良好的效果和可解释性, 已成为一种广泛应用的图分析方法. 然而现有主流的深度学习框架(如TensorFlow、PyTorch等)没有为图神经网络计算提供高效的存储支持和图上的消息传递支持, 这限制了图神经网络算法在大规模图数据上的应用. 目前已有诸多工作针对图结构的数据特点和图神经网络的计算特点, 探索了大规模图神经网络系统的设计和实现方案. 首先对图神经网络的发展进行简要概述, 总结了设计图神经网络系统需要面对的挑战; 随后对目前图神经网络系统的工作进行介绍, 从系统架构、编程模型、消息传递优化、图分区策略、通信优化等多个方面对系统进行分析; 最后使用部分已开源的图神经网络系统进行实验评估, 从精确度、性能、扩展性等多个方面验证这些系统的有效性. |
| 关键词: 图神经网络 大规模图数据 分布式系统 深度学习 反向传播 |
| DOI：10.13328/j.cnki.jos.006311 |
| 分类号: |
| 基金项目:国家重点研发计划 (2018YFB1003400); 国家自然科学基金(61672141, 62072082); 中央高校基本科研业务费(N181605017, N181604016) |
|
| Survey on Large-scale Graph Neural Network Systems |
|
ZHAO Gang, WANG Qian-Ge, YAO Feng, ZHANG Yan-Feng, YU Ge
|
|
School of Computer Science and Engineering, Northeastern University, Shenyang 110169, China
|
| Abstract: |
| Graph neural network (GNN) is used to process graph structure data based on deep learning techniques. It combines graph propagation operations with deep learning algorithms to fully utilize graph structure information and vertex features in the learning process. GNNs have been widely used in a range of applications, such as node classification, graph classification, and link prediction, showing promised effectiveness and interpretability. However, the existing deep learning frameworks (such as TensorFlow and PyTorch) do not provide efficient storage support and message passing support for GNN’s training, which limits its usage on large-scale graph data. At present, a number of large-scale GNN systems have been designed by considering the data characteristics of graph structure and the computational characteristics of GNNs. This study first briefly reviews the GNNs, and summarizes the challenges that need to be faced in designing GNN systems. Then, the existing work on GNN training systems is reviewed, and these systems are analyzed from multiple aspects such as system architecture, programming model, message passing optimization, graph partitioning strategy and communication optimization. Finally, several open source GNN systems are chosen for experimental evaluation to compare these systems in terms of accuracy, efficiency, and scalability. |
| Key words: graph neural network (GNN) large-scale graph data distributed system deep learning back propagation |
