Our new paper TabuLa: Harnessing Language Models for Tabular Data Synthesis is on arxiv now! The code is published here. Tabula improves tabular data synthesis by leveraging language model structures without the burden of pre-trained model weights. It offers a faster training process by preprocessing tabular data to shorten token sequence, which sharply reducing training time while consistently delivering higher-quality synthetic data. Its training time is longer than CTAB-GAN+, but the synthetic data fidelity is amazing! It also works for high-dimentional categorical columns!
The CTAB-GAN+ code is released. CTAB-GAN+ updates the CTAB-GAN with new losses (i.e., WGAN+GP) and new feature engineering (i.e., general transform), the training is more stable and efficient. The problem type supports Classification and Regression dataset. You can also indicate the problem_type as None in CTAB-GAN+ code.
This is the official git paper CTAB-GAN: Effective Table Data Synthesizing. The paper is published on Asian Conference on Machine Learning (ACML 2021), please check our pdf on PMLR website for our newest version of paper, it adds more content on time consumption analysis of training CTAB-GAN. If you have any question, please contact z.zhao-8@tudelft.nl for more information.
The required package version
numpy==1.21.0
torch==1.9.1
pandas==1.2.4
sklearn==0.24.1
dython==0.6.4.post1
scipy==1.4.1The sklean package in newer version has updated its function for sklearn.mixture.BayesianGaussianMixture. Therefore, user should use this proposed sklearn version to successfully run the code!
Experiment_Script_Adult.ipynb is an example notebook for training CTAB-GAN with Adult dataset. The dataset is alread under Real_Datasets folder.
The evaluation code is also provided.
If your dataset has large number of column, you may encounter the problem that our currnet code cannot encode all of your data since CTAB-GAN will wrap the encoded data into an image-like format. What you can do is changing the line 341 and 348 in model/synthesizer/ctabgan_synthesizer.py. The number in the slide list
sides = [4, 8, 16, 24, 32]is the side size of image. You can enlarge the list to [4, 8, 16, 24, 32, 64] or [4, 8, 16, 24, 32, 64, 128] for accepting larger dataset.
To cite this paper, you could use this bibtex
@InProceedings{zhao21,
title = {CTAB-GAN: Effective Table Data Synthesizing},
author = {Zhao, Zilong and Kunar, Aditya and Birke, Robert and Chen, Lydia Y.},
booktitle = {Proceedings of The 13th Asian Conference on Machine Learning},
pages = {97--112},
year = {2021},
editor = {Balasubramanian, Vineeth N. and Tsang, Ivor},
volume = {157},
series = {Proceedings of Machine Learning Research},
month = {17--19 Nov},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v157/zhao21a/zhao21a.pdf},
url = {https://proceedings.mlr.press/v157/zhao21a.html}
}