We consider three sets of datasets in this paper.

• Traffic dataset: The python library torch-spatiotemporal provides two preprocessed datasets (PEMS-BAY, METR-LA).

• Large-scale dataset: The downloaded datasets are suggested to be stored in ./data/pvus and ./data/cer, respectively. ./data/pre_pvus.py and ./data/pre_cer.py are used to preprocess these two datasets.

  • PV-US: This dataset can be downloaded from https://www.nrel.gov/grid/solar-power-data.html.
  • CER-E: This dataset can be downloaded from https://www.ucd.ie/issda/data/commissionforenergyregulationcer/.

• Traffic4cast dataset: This dataset can be downloaded from this project.

See requirements.txt for the list of packages.

• Traffic dataset: main_traffic.py is used to train the proposed OPCR on traffic dataset.

"""
gpu: device id
dataset-name: "bay" or "la"
missing-type: "point" or "spatial"
missing-rate: [0.0, 1.0]
seed: random seed
"""
python main_traffic.py --gpu 0 --dataset-name bay --missing-type point --missing-rate 0.95 --seed 0

• Large-scale dataset: main_large.py is used to train the proposed OPCR on large-scale dataset.

"""
gpu: device id
dataset-name: "pvus" or "cer"
missing-type: "point" or "spatial"
missing-rate: [0.0, 1.0]
seed: random seed
"""
python main_large.py --gpu 0 --dataset-name pvus --missing-type point --missing-rate 0.95 --seed 0

• Traffic4cast dataset: main_t4c_con.py and main_t4c_seg.py are used to train the proposed OPCR on t4c dataset.

"""
gpu: device id
city: "london" or "madrid" or "melbourne"
"""
# congestion classification task
python main_t4c_con.py --gpu 0 --city london
# travel time prediction task
python main_t4c_seg.py --gpu 0 --city london

@inproceedings{denglearning,
  title={Learning from Highly Sparse Spatio-temporal Data},
  author={Deng, Leyan and Wu, Chenwang and Lian, Defu and Chen, Enhong},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}
}

Our experiments on the traffic dataset and strategy for injecting missing data are based on the implementations of SPIN. We gratefully thanks for their contribution.