Multi-ATGCN: A Multi-View Graph Neural Network-based Framework for Citywide Crowd Inflow Forecasting

This is an original PyTorch implementation of Multi-ATGCN in the following working paper:

Multi-ATGCN: A Multi-View Graph Neural Network-based Framework for Citywide Crowd Inflow Forecasting

Environment

We use the torch == 1.10.2 and Python 3.6.11 for implementation.

We follow the framework of LibCity to prepare data and run the model. See more details in the requirement.txt and the LibCity document.

Data Preparation

Data files for Washington, D.C. and Baltimore City are available at the raw_data/ folder. Please extract them to the current fold and you will get a set of atomic files following the LibCity Unified Data Structures:

filename	content	example
xxx.geo	Store geographic entity attribute information.	geo_id, type, coordinates
xxx.rel	Store the relationship information between entities, i.e. the adjacency matrix.	rel_id, type, origin_id, destination_id, link_weight
xxx.dyna	Store hourly crowd flow information.	dyna_id, type, time, entity_id, Visits
xxx.ext	Store external time-varying information, such as weather, holidays, etc.	ext_id, time, properties[...]
xxx.static	Store external static information, such as socioeconomics, POIs, demographics.	geo_id, properties[...]
xxx.gbst	Store mean and std for each geo unit before the group-based z-score.	geo_id, mean, std
config.json	Used to supplement the description of the above table information.

The .dyna files are retrieved from SafeGraph using the Weekly Places Patterns Dataset. Run codes at ./data_prepare to prepare the data and transfer it to the required format if you own the access to the raw Weekly Places Patterns Dataset.

The data statistics of two datasets are as followed:

	Washington, D.C.	Baltimore City (and surrounding counties)
Date Range	01/01/2019 - 05/31/2019	01/01/2019 - 05/31/2019
# Nodes	237	403
# Samples	858,888	1,460,472
Sample Rate	1 hour	1 hour
Input length	24 hours	24 hours
Output length	3 hours, 6 hours, 12 hours, 24 hours	3 hours, 6 hours, 12 hours, 24 hours
Mean of crowd flow	30.169	14.41
St.d. of crowd flow	84.023	29.3

Code Structure

For easy comparison among different models, the code and data formats follow the framework proposed by LibCity:

• The code for Multi-ATGCN is located at ./libcity/model/traffic_flow_prediction/MultiATGCN.py.
• The code for dataset preprocessing is located at ./libcity/data/dataset/dataset_subclass/mth_dataset.py.
• The configuration for the model is located at ./libcity/config/model/traffic_state_pred/MultiATGCN.json.
• The user-defined configuration with the highest priority is located at ./config_user.json.

❗ You can also directly copy the data and our model to the LibCity environment and run. However, I suggest using the repository here since some changes are made compared with the original LibCity:

• A new data format for static variables is added.
• A new dataset class, the mth_dataset, is added, to support multi-head temporal fusion across all models.
• Support group-based normalization in model evaluation.
• Support separate inclusion of time-varying external variables and time-varying calendar variables.
• Configurations of model, data, and executor are changed accordingly to fit our dataset.

Model Training

The script run_model.py is used for training and evaluating the main model:

# DC
python run_model.py --task traffic_state_pred --dataset 201901010601_DC_SG_CTractFIPS_Hourly_Single_GP

# Baltimore
python run_model.py --task traffic_state_pred --dataset 201901010601_BM_SG_CTractFIPS_Hourly_Single_GP

The script run_model_parameters.py is used for parameter study and ablation study. Change the parameters you are interested and run:

# DC
python run_model_parameters.py --task traffic_state_pred --dataset 201901010601_DC_SG_CTractFIPS_Hourly_Single_GP

# Baltimore
python run_model_parameters.py --task traffic_state_pred --dataset 201901010601_BM_SG_CTractFIPS_Hourly_Single_GP

If you are using Google Colab, we also provide a notebook named Colab_run_model.ipynb to execute in Colab environment. Clone all files to your drive and execute the code blocks successively.

Results

See details in our paper.

Top and last three census tracts (Baltimore)

Top and last three census tracts' forecasting results

Top and last three census tracts (DC)

Top and last three census tracts' forecasting results