MargCTGAN

The X-axis represents the size of the training dataset, with "all" indicating the full dataset size. Real data (reference) corresponds to the metrics directly measured on the real (train vs. test) data, serving as the reference (oracle score) for optimal performance.

This repository contains the implementation for "MargCTGAN: A ``Marginally'' Better CTGAN for the Low Sample Regime" Published at GCPR 2023. This paper was also presented at ICML 2023 Deploying Generative AI Workshop.

Authors: Tejumade Afonja, Dingfan Chen, and Mario Fritz

Contact: Tejumade Afonja (tejumade.afonja@cispa.de)

Requirements

This implementation is based on PyTorch (tested for version 1.7.0). Please refer to requirements.txt for the other required packages and version.

Datasets

The implementation supports the following datasets:

Adult, Census, News, and Texas.

You can download the data by running the notebooks in data/<dataset_name>/<dataset_name>-download.ipynb.

Running Experiments

API (Run experiments using the default configurations).

Change to the synthesizers/<name_of_model> directory and run the code snippet below (after updating the placeholders).

python <name_of_model_main_script>.py \
    -name <name_of_experiment> \
    -data <dataset_name> \
    -ep <number_of_epoch> \
    -s <seed> \
    --train "True" \
    --evaluate "True" \
    --sample "True" \
    --subset_size <size_of_real_dataset> \
    --sample_size <size_of_synthetic_dataset> \
    --eval_retries <how_many_times_to_rerun_evaluation>

Citation

@inproceedings{afonja2023margctgan,
  title={MargCTGAN: A “Marginally” Better CTGAN for the Low Sample Regime},
  author={Afonja, Tejumade and Chen, Dingfan and Fritz, Mario},
  booktitle={DAGM German Conference on Pattern Recognition},
  pages={524--537},
  year={2023},
  organization={Springer}
}

Acknowledgements

Our implementation uses the source code from the following repositories:

Margctgan

Install / Use

README