摘要
How to select a suitable spatial weighting scheme for convolutional graph neural networks (ConvGNNs) is challenging. In this study, we propose a ConvGNN, termed learnable graph convolutional (LGC) network, which learns spatial weightings between a road and its k -hop neighbours as learnable parameters in the spatial convolutional operator. A dynamic LGC (DLGC) network is further proposed to learn the dynamics of spatial weightings by explicitly considering the tem-poral correlations of spatial weightings at different times of the day. A multi-temporal DLGC (MTDLGC) network is developed for forecast-ing traffic variables in road networks. Results of case study suggest that the MT-DLGC network can achieve higher prediction accuracy than other state-of-the-art baselines. Both LGC and DLGC networks can be used as general spatial weighting schemes for baselines with better forecasting performance than existing spatial weight-ing schemes, e.g., graph attention. The source code of this study is available publicly at https://github.com/Mayaohong/MTDLGC.
学科与关键词
- Graph neural network for traffic forecasting: A surveyJiang, Weiwei*; Luo, JiayunEXPERT SYSTEMS WITH APPLICATIONS, 2022, 207: 117921.被引 433
- Spatial-Temporal Convolutional Graph Attention Networks for Citywide Traffic Flow ForecastingZhang, Xiyue; Huang, Chao; Xu, Yong*; Xia, Lianghao29th ACM International Conference on Information and Knowledge Management (CIKM), ELECTR NETWORK, 2020-10-19 To 2020-10-23.被引 38
- Multihead Attention-Based Multiscale Graph Convolution Network for ITS Traffic ForecastingDeng, Zhiyuan; Hou, Yue*; Jolfaei, Alireza; Zhou, Wei; Farivar, Faezeh; Haghighi, Mohammad SayadIEEE Systems Journal, 2024, 18(2): 836-847.
