DSTGCN is a graph-based neural network that predicts the risk of traffic accidents in the future.

Please refer to our Neurocomputing 2021 paper “Deep Spatio-Temporal Graph Convolutional Network for Traffic Accident Prediction” for more details.

The descriptions of principal files in this project are introduced as follows:

• model/
  • spatial_layer.py: the Spatial Convolutional Layer
  • spatial_temporal_layer.py: the Spatial-Temporal Convolutional Layer
  • fully_connected.py: fully connected network for the Embedding layer
  • DSTGCN.py: the Deep Spatio-Temporal Graph Convolutional Network
• preprocessing_data/
  • process_xxx.py: process the raw 'xxx' dataset
  • generate_data.py: generate the dataset for our model
• transform_coord/: convert the coordinate, which could be obtained from here.
• train/
  • train_model.py and train_main.py: train models
• test/
  • test_main.py: evaluate models
• utils/: utility files (e.g. data loader, metrics calculation, loss function, configurations)
• original_data/: original datasets. Due to the data privacy, we do not provide the original data. But, you could use the preprocess codes in preprocessing_data/ folder to generate your own datasets and use them to train the model.
• data/: processed datasets. We provide a sampled dataset here. You can download it and then put the data files in this folder.
• saves/ and runs/: folders to save models and outputs of tensorboardX, respectively
• results/: folders to save the evaluation metrics for models.

• accident data format: records of traffic accidents.

• poi data format: records of pois.

• road data (beijing_roadnet.gpickle) format: networkx.classes.graph.Graph, road network structure that records the connectivity of road segments. You can download the original data from here, and then run preprocessing_data/process_beijing_road_net.py to get the preprocessed file.

• speed data (all_grids_speed.h5) format: DataFrame, containing the traffic speed of each grid

• weather data format (weather.h5) format: records of the weather condition. You can run preprocessing_data/process_weather.py to get the preprocessed file.

• edge.h5 data format: records of spatial features (poi and road segment features), which is a combination of preprocessed poi data and road data.

Please refer to our paper for more details of parameter settings. Hyperparameters could be found in utils/config.json and you can adjust them when running the model.

• Training: after setting the parameters, run python train_main.py to train models.
• Testing: run python test_main.py to evaluate models based on the path of saved models.

• PyTorch == 1.8.1
• dgl == 0.7.0
• tqdm
• numpy
• pandas
• sklearn
• tensorboardX

Please consider citing the following paper when using our data or code.

@article{DBLP:journals/ijon/YuDHSHL21,
  author    = {Le Yu and
               Bowen Du and
               Xiao Hu and
               Leilei Sun and
               Liangzhe Han and
               Weifeng Lv},
  title     = {Deep spatio-temporal graph convolutional network for traffic accident
               prediction},
  journal   = {Neurocomputing},
  volume    = {423},
  pages     = {135--147},
  year      = {2021}
}