Baselines and benchmarks for spatio-temporal forecasting problems in transportation
This repository contains source code for reproducing the results in:
It consists of code for computing the "HA" and "HA+LR" baselines described in the paper, as well as code for preparing the experimental setups (i.e., train/val/test splits, forecasting horizons, evaluation metrics, etc.) for the following 9 publicly-available datasets:
| Name | Type | Timespan | Time granularity | Train/val/test split | Source |
|---|---|---|---|---|---|
| PeMSD7(M) - California | traffic speeds | 01/04/2016 - 30/06/2016 | 5 minutes | 34/5/5 days | Yu et al., 2018 [1] |
| Urban1 - South Korea | traffic speeds | 01/04/2018 - 30/04/2018 | 5 minutes | 70/10/20 % | Lee and Rhee, 2022 [2] |
| NYC Citi Bike - New York | pickups and dropoffs | 01/04/2016 - 01/04/2016 | 30 minutes | 63/14/14 days | Ye et al., 2021 [3] |
| PeMSD4 - California | traffic volumes | 01/01/2018 - 28/02/2018 | 5 minutes | 60/20/20 % | Choi et al., 2022 [4] |
| SZ-taxi - Shenzhen | traffic speeds | 01/01/2015 - 31/01/2015 | 15 minutes | 80/-/20 % | Zhao et al., 2021 [5] |
| METR-LA - Los Angeles | traffic speeds | 01/03/2012 - 30/06/2012 | 5 minutes | 70/10/20 % | Li et al., 2018 [6] |
| PEMS-BAY - California | traffic speeds | 01/01/2017 - 31/05/2017 | 5 minutes | 70/10/20 % | Li et al., 2018 [6] |
| NYC Citi Bike - New York | in- and out-flows | 01/07/2017 - 30/09/2017 | 1 hour | 80/10/10 % | Xia et al., 2021 [7] |
| Seattle loop data - Seattle | traffic speeds | 01/11/2015 - 31/12/2015 | 5 minutes | 56/-/5 days | Yang et al., 2021 [8] |
The goal is to facilitate the comparison between different spatio-temporal forecasting approaches by providing multiple well-defined reference benchmarks.
The repo does not include the actual data, but the table below provides links to where the data can be downloaded:
| Name | Download link | Jupyter notebook with baseline |
|---|---|---|
| PeMSD7(M) - California | Link (Mirror) | STGCN - Historical average baselines-FINAL.ipynb |
| Urban1 - South Korea | Link (Mirror) | DDP-GCN - Historical average baselines-FINAL.ipynb |
| NYC Citi Bike - New York | Link (Mirror) | CGCDemand - Historical average baselines-seq2seq.ipynb |
| PeMSD4 - California | Link (Mirror) | STG-NCDE - Historical average baselines-FINAL-seq2seq.ipynb |
| SZ-taxi - Shenzhen | Link (Mirror) | T-GCN - Mobility baselines.ipynb |
| METR-LA - Los Angeles | Link (Mirror) | DCRNN - Mobility baselines - METR-LA.ipynb |
| PEMS-BAY - California | Link (Mirror) | DCRNN - Mobility baselines - PEMS-BAY.ipynb |
| NYC Citi Bike - New York | Link (Mirror) | 3D-DGCN - Mobility baselines-FINAL.ipynb |
| Seattle loop data - Seattle | Link (Mirror) | TransPAI - Mobility baselines.ipynb |
The file mobility_detrender.py contains the code for computing the historical averages (HA).
References:
- [1] B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,” in 27th International Joint Conference on Artificial Intelligence (IJCAI-18), 2018.
- [2] K. Lee and W. Rhee, “DDP-GCN: Multi-graph convolutional network for spatiotemporal traffic forecasting,” Transportation Research Part C: Emerging Technologies, vol. 134, p. 103466, 2022.
- [3] J. Ye, L. Sun, B. Du, Y. Fu, and H. Xiong, “Coupled layer-wise graph convolution for transportation demand prediction,” in Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21), 2021.
- [4] J. Choi, H. Choi, J. Hwang, and N. Park, “Graph neural controlled differential equations for traffic forecasting,” in Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22), 2022.
- [5] L. Zhao, Y. Song, C. Zhang, Y. Liu, P. Wang, T. Lin, M. Deng, and H. Li, “T-GCN: A temporal graph convolutional network for traffic prediction,” IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 9, pp. 3848–3858, 2019.
- [6] Y. Li, R. Yu, C. Shahabi, and Y. Liu, “Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,” in International Conference on Learning Representations (ICLR-18), 2018.
- [7] T. Xia, J. Lin, Y. Li, J. Feng, P. Hui, F. Sun, D. Guo, and D. Jin, “3DGCN: 3-dimensional dynamic graph convolutional network for citywide crowd flow prediction,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 15, no. 6, pp. 1–21, 2021.
- [8] J.-M. Yang, Z.-R. Peng, and L. Lin, “Real-time spatiotemporal prediction and imputation of traffic status based on lstm and graph laplacian regularized matrix factorization,” Transportation Research Part C: Emerging Technologies, vol. 129, p. 103228, 2021.