City-wide Origin-destination Matrix Generation via Cascaded Graph Denoising Diffusion.

This repo contains the codes and data for our submitted KDD'25 applied data science track paper under review.

The models are trained on RTX3090-24G.

Environment

• python == 3.8.13
• torch == 1.12.1+cu113

Files

• code # python scripts
  • utils # tool codes
    • metrics.py # all metrics
    • MyLogger.py # A logger for logging experimental information
    • procedure.py # pipeline function
    • tool.py # simple tool functions
  • data_load.py # load data
  • eval.py #
  • main.py # entry
  • model.py # models
  • train.py # training scripts
• data # datasets
• exp # experimental information
  • config # configurations
  • logs # losses and evaluation results
  • results # generations
  • running #
  • runs # for tensorboard
  • weightes # trained model parameters

Usage

• The experimental configuration can be adapted in exp/config/xxx.json
• In the config file, adjust the exp_name to record meta information for different experiments.
• In the config file, adjust src_cities and tar_cities to select the cities for training and testing. The names of the cities need to be consistent with the names of the subdirectories in the data directory
• In main.py, modify the selected config file.
• Trained models have been saved in exp/weights. Adjust exp_name to load them.

training

• set topo_train to 1 means training the topology diffusion model
• set flow_train to 1 means training the flow diffusion model
• set T_mode to INIT means training the topology diffusion model from scratch
• set F_mode to INIT means training the flow diffusion model from scratch
• Extra setting
  • set teacher_force to 1 means training the flow diffusion models in collaborative mode
  • set mem_need to check GPU memory, at leat 23000

testing

• set topo_train and flow_train to 0 to skip the training process
• set T_mode and F_mode to LOAD to load existing trained models