Graph Convolutional Adversarial Networks for Spatio-Temporal Anomaly Detection

Two datasets are available at Google Drive. If you use the data, please cite the following paper.

@ARTICLE{9669110,
  author={Deng, Leyan and Lian, Defu and Huang, Zhenya and Chen, Enhong},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={Graph Convolutional Adversarial Networks for Spatiotemporal Anomaly Detection}, 
  year={2022},
  volume={},
  number={},
  pages={1-13},
  doi={10.1109/TNNLS.2021.3136171}}

PeMS dataset (bay) is collected from PeMS and NYC dataset (nyc) is provided by Detecting Collective Anomalies from Multiple Spatio- Temporal Datasets across Different Domains. Each dataset consists of the following five datasets:

• data.npy
• node_subgraph.npy
• node_adjacent.txt
• time_features.txt
• node_dist.txt

where the data.npy is the traffic data in Bay area or New York City; node_subgraph.npy represents the adjacency matrix of the subgraph of each node; node_adjacent.txt represents all nodes in the subgraph of each node; time_features.txt represents the time feature of each time slots; node_dist.txt represents the distance between nodes.

We also provide the information of the selected sensors in our paper, the file is vds_info.csv.

We have updated the ground truth of PeMS datasets. Note that we didn't delete the CHP incidents with duration<=0 since we extended the end time of each incident by 1 hour to include the impact of the traffic accidents. The ground truth of NYC dataset is provided by the authors of paper Detecting Collective Anomalies from Multiple Spatio-Temporal Datasets across Different Domains.

If you need anomaly labels for other times, please refer to CHP incident and LCS Report. I hope they will be helpful to you.

If you have any question about the code or the paper, please contact me by email ([email protected]).