This is a PyTorch implementation of Deep Kalman Filtering Network(DKFN) in the following poster paper:
Fanglan Chen, Zhiqian Chen, Subhodip Biswas, Shuo Lei, Naren Ramakrishnan, and Chang-Tien Lu, Graph Convolutional Networks with Kalman Filtering for Traffic Prediction, SIGSPATIAL 2020.

• torch==1.0.1
• numpy>=1.16.5
• pandas>=0.25.1
• argparse
• time
• math

The DKFN model is tested on two real-world traffic speed datasets, METR-LA data and Seattle-Loop data. The METR-LA data is shared in the METR_LA_Dataset folder. The Seattle-Loop data is acquired from this repository. You can go to the original link to download the 2015 year-round data or assess a 3-month (Jun-Aug) via Google Drive.

To run the code, you need to download the traffic speed data and graph adjacency matrix and put them in the corresponding "_Dataset" folder.

To train and test DKFN with specific parameters, you can optionally change the main.py file and here are commands for training/testing the DKFN model on METR-LA and Seattle-Loop respectively.

# METR-LA
python main.py -dataset=metr-la -model=dkfn

# Seattle-Loop
python main.py -dataset=seattle-loop -model=dkfn

For baselines including RNN, LSTM, and GCLSTM, you can change the model name to rnn, lstm, or gclstm, respectively.

If you find this repository useful in your own research, please cite the following poster paper:

@inproceedings{chen2020graph,
  title={Graph Convolutional Networks with Kalman Filtering for Traffic Prediction},
  author={Chen, Fanglan and Chen, Zhiqian and Biswas, Subhodip and Lei, Shuo and Ramakrishnan, Naren and Lu, Chang-Tien},
  booktitle={Proceedings of the 28th International Conference on Advances in Geographic Information Systems},
  pages={135--138},
  year={2020}
}