1、arXiv:1707.01926v1[cs.LG]6Jul20171Graph Convolutional Recurrent Neural Network:Data-Driven Traffic ForecastingYaguang Li Rose Yu Cyrus Shahabi Yan LiuDepartment of Computer Science, University of Southern California{yaguang, qiyu, shahabi, yanliu.cs}@usc.eduAbstractSpatiot
2、emporal forecasting has significant implications in sustainability, transportationand health-care domain. Traffic forecasting is one canonical example of such learningtask. This task is challenging due to (1) non-linear temporal dynamics with changing roadconditions, (2) com
3、plex spatial dependencies on road networks topology and (3) inherentdifficulty of long-term time series forecasting. To address these challenges, we proposeGraph Convolutional Recurrent Neural Network to incorporate both spatial and temporaldependency in traffic flow. We furth
4、er integrate the encoder-decoder framework and scheduledsampling to improve long-term forecasting. When evaluated on real-world road networktraffic data, our approach can accurately capture spatiotemporal correlations and consistentlyoutperforms state-of-the-art baselines b
5、y 12% - 15%.IntroductionSpatiotemporal forecasting is a crucial task for a learning system that operates in a dynamicenvironment. Accurate spatiotemporal forecasting has a wide range of applications ranging fromvideo compression and understanding, as to energy and smart g
6、rid management, economics andfinance, to environmental and health care. In this paper, we study one example of spatiotemporalforecasting task: traffic forecasting, the core component of the intelligent transportation systems.We believe our approach is not limited to transpor
7、tation, and is readily applicable to otherdomains as well.The goal of traffic forecasting is to predict the future speeds of a sensor network usingprevious traffic speeds as well as the underlying road networks structure. This task is challengingmainly due to the complex spat
8、ial and temporal dependencies. On one hand, traffic time seriesdemonstrate strong temporal dynamic
