TVGCN: Time-variant graph convolutional network for traffic forecasting

被引:29
作者
Wang, Yuhu [1 ,2 ]
Fang, Shen [1 ,2 ]
Zhang, Chunxia [3 ]
Xiang, Shiming [1 ,2 ]
Pan, Chunhong [1 ]
机构
[1] Chinese Acad Sci, Inst Automat, Natl Lab Pattern Recognit, Beijing 100190, Peoples R China
[2] Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing 101408, Peoples R China
[3] Beijing Inst Technol, Sch Comp Sci & Technol, Beijing 100081, Peoples R China
基金
中国国家自然科学基金;
关键词
Spatial-temporal correlation; Graph convolutional network; Traffic forecasting; FLOW;
D O I
10.1016/j.neucom.2021.11.006
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Traffic forecasting is a very challenging task due to the complicated and dynamic spatial-temporal correlations between traffic nodes. Most existing methods measure the spatial correlations by defining physical or virtual graphs with distance or similarity measurement, which is constructed with stable edge connections by some prior knowledge. However, the use of such graphs with stable edge connections limits the variations of spatial correlations between traffic nodes at different times, which can not capture the hidden dynamic patterns of traffic graphs. This paper proposes a Time-Variant Graph Convolutional Network (TVGCN) to overcome this limitation. Architecturally, a time-variant spatial convolutional module (TV-SCM) is developed on two graphs without any prior knowledge. One graph is learned to capture the stable spatial correlations of the traffic graph, while the other graph is evolved to model dynamic spatial correlations at different times. Such two graphs are combined hierarchically together under the framework of graph convolutional network (GCN). Moreover, a gated multi-scale temporal convolutional module (GMS-TCM) is designed to extract long-range temporal dependencies within traffic nodes, which are further supplied to the TV-SCM to mutually explore the spatial correlations between traffic nodes. Extensive experiments conducted on three real-world traffic datasets indicate the effectiveness and superiority of our proposed approach. (c) 2021 Elsevier B.V. All rights reserved.
引用
收藏
页码:118 / 129
页数:12
相关论文
共 49 条
[1]  
[Anonymous], 2020, AAAI CONF ARTIF INTE, DOI 10.1609/aaai.v34i01.5477
[2]  
Bai L, 2020, ADV NEUR IN, V33
[3]  
Bai S., ARXIV PREPRINT ARXIV
[4]  
Bruna J., ARXIV PREPRINT ARXIV
[5]   Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting [J].
Cai, Ling ;
Janowicz, Krzysztof ;
Mai, Gengchen ;
Yan, Bo ;
Zhu, Rui .
TRANSACTIONS IN GIS, 2020, 24 (03) :736-755
[6]   Local-Aggregation Graph Networks [J].
Chang, Jianlong ;
Wang, Lingfeng ;
Meng, Gaofeng ;
Zhang, Qi ;
Xiang, Shiming ;
Pan, Chunhong .
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2020, 42 (11) :2874-2886
[7]  
Chung Junyoung, 2014, EMPIRICAL EVALUATION
[8]  
Defferrard M, 2016, ADV NEUR IN, V29
[9]  
Drucker H, 1997, ADV NEUR IN, V9, P155
[10]  
Fang S, 2019, PROCEEDINGS OF THE TWENTY-EIGHTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE, P2286