An implementation of 3D-DGCN: 3D Dynamic Graph Convolutional Network for Crowd Flow Prediction. The paper is submitted to ICDE 2020.
Our datasets is from the NYC Bike. We obtained road network and 9 categories of PoIs from OpenStreatMap. In two datasets, the city is devided into 128 regular regions and 82 irregular regions, respectively. Each dataset contains 1448 time intervals, spanning from Jul. 1st to Sept. 30th, 2017.
- python 2.7
- PyTorch 0.2.0
- NumPy
- JSON
- /flow
- flow_bike_nyc_regular.json contains the regular dataset
- flow_bike_nyc_irregular.json contains the irregular dataset
- /poi (* = regular or irregular)
- *_feature.npy is the features of all regions
$\bm{F}$ - *_idx.npy is the index of labeled regions
$\mathcal{V}_L$ - *_label.npy is the labels of all regions
$\bm{\Omega}$ - *_weight.npy is the normalization term
$frac{1}{|\mathcal{V}_L^{\Omega_i}|}$
- *_feature.npy is the features of all regions
- /path
- regular_path.npy contains the historical flow paths for the regular dataset
- irregular_path.npy contains the historical flow paths for the irregular dataset
- /main.py run this file to get the results of 3D-DGCN in our paper
- /gcn.py is the implementation of our neural network
- /dataset.py loads the dataset
python main.py