区域-道路时空图网络:一种基于图神经网络的流量预测模型  

Region to Road Spatial-temporal Graph Networks:a Framework for Traffic Forecasting

作　　者：金高铭 刘安[1] 孙玉娥 于金刚[3] JIN Gao-ming;LIU An;SUN Yu-e;YU Jin-gang(School of Computer Science and Technology,Soochow University,Suzhou 215006,China;School of Rail Transportation,Soochow University,Suzhou 215137,China;Shenyang Institute of Computing Technology,Chinese Academy of Sciences University,Shenyang 110168,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]苏州大学轨道交通学院,江苏苏州215137 [3]中国科学院沈阳计算技术研究所,沈阳110168

出　　处：《小型微型计算机系统》2022年第9期1801-1807,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61572336)资助.

摘　　要：交通预测在城市规划中具有重要意义.由于交通流具有复杂的时空相关性,交通预测的任务一直面临着许多挑战.现有的方法通常利用上层区域图上的观测特征序列来学习和评价交通状况,而忽略了具有丰富语义的底层路网.为了克服这一缺点,本文将空间数据建模为观测图和隐藏图,并提出了一种新的交通预测框架:区域-道路时空图网络(R2RSTGN).为了学习道路网络和观测到的交通信息之间隐藏的相互作用,本文同时考虑区域和细粒度道路并提出了一种压缩注意力机制,可以为预测提供解释结果.在两个真实数据集上的实验结果验证了该框架的有效性.Traffic prediction has shown great importance in urban planning.However,the task of traffic prediction always involves many challenges due to complex spatial-temporal correlations in traffic flow.Existing methods usually consider the aggregated feature sequence as an observation on the region graph to learn and evaluate traffic conditions,ignoring the underlying road network with rich semantics.To overcome this shortcoming,we model in this paper the spatial data structure as observed and hidden graphs and propose a novel framework,named Region to Road Spatial-Temporal Graph Network(R2RSTGN)for traffic prediction.To learn the hidden interplay between the road network and the observed traffic information,we propose a compressing attention mechanism,which can provide an explainable result for the prediction.To the best of our knowledge,this is the first work that considers both regions and fine-grained roads.Experimental results on two real-world traffic prediction tasks verify the effectiveness of the proposed framework.

关 键 词：图模型 交通预测 神经网络 序列模型 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]