Causal-TGAN

This is the code for the project Causal-TGAN: Causally-Aware Tabular Data Generative Adversarial Networks

Citation

If you use this code for your research, please cite:

@inproceedings{
wen2022causaltgan,
title={Causal-{TGAN}: Modeling Tabular Data Using Causally-Aware {GAN}},
author={Bingyang Wen and Yupeng Cao and Fan Yang and Koduvayur Subbalakshmi and Rajarathnam Chandramouli},
booktitle={ICLR Workshop on Deep Generative Models for Highly Structured Data},
year={2022},
url={https://openreview.net/forum?id=BEhxCh4dvW5}
}

Prerequisties

The project is built on python3.6 with the pytorch version of 1.9.0.

To run this project, please add a new python environment path to this project. For example if your cloned repository reside in /home/username/CausalTGAN, then one way to do this is export PYTHONPATH="/home/username" from command line or add it to your ~/.bashrc.

To prepare a dataset for training Causal-TGAN

• An example dataset is illustrated in ./data/real_world/adult.
  

• Instructions of adding the customized dataset is following:

  • Step 1: Add your dataset name and name of column information into helper/constant.py. For example:

   DATASETS = ['asia', 'alarm', 'child', 'insurance', 'your dataset name']
   YOUR_DATASET_NAME = ['column_1', 'column_2', 'column_3']

  • Step 2: Check your dataset type (Continuous, Discrete, Mix) and add the dataset type information into check_BN_datatype() function in helper/utils.py. For example, if your dataset is Continuous type dataset, you can modify the code as following:

   def check_BN_datatype(data_name):
   ...
   
         if data_name in ['your_dataset_name']:
             return 'continuous'
         
   ...

  • Step 3 (Optional): If customized dataset is Continuous type, please declare Discrete Column in get_discrete_cols() function in helper/utils.py. For example,

   def get_discrete_cols(data, data_name):
   ...
   
     if data_name == 'your_dataset_name':
         discrete_cols = Discrete_Column_Name
     
   ...

  • Step 4: Prepare your causal graph by using Tetrad. You can go to https://www.ccd.pitt.edu/tools/ and click [Tetrad Latest Executable] to download Tetrad. We have created a tutorial on how to use Tetrad to generate causal graphs. See in Tetrad_Tutorial.pdf
    After this steps, you should have a 'graph_tetrad.txt' file.

  • Step 5: Use helper/graphFromTetrad.py to convert Tetrad graph to common graph data and save it as 'graph.txt'.

• Now, you have completed the preparation steps for the dataset. Your dataset folder structure should look like the following:

|-data
|  |-customized_dataset
|     |-graph.txt
|     |-train.csv
|     |-test.csv

Then, you can use the load_data_graph(args.data_name) function in train.py to load your customized datasets and train model.

To train a Causal-TGAN

python train.py. The training details such as dataset and epochs are set inside train.py

To sample from Causal-TGAN

python sampling.py. The details such as number of samples to generate and causal-TGAN path are set inside sampling.py