# Modeling Tabular data using Conditional GAN

**Authors:** Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, Kalyan Veeramachaneni  
**arXiv:** [1907.00503](https://arxiv.org/abs/1907.00503)  
**Date:** 2019-07-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1907.00503v1/

## Key points

- The paper addresses the task of modeling the probability distribution of rows in tabular data and generating realistic synthetic data.
- Existing statistical and deep neural network models struggle to accurately model this type of data.
- The authors propose TGAN, a method that utilizes a conditional generative adversarial network to effectively model the complex distribution of tabular data.
- TGAN captures both the discrete-continuous nature and potential imbalance in the columns.
- The authors create a benchmark consisting of 7 simulated datasets and 8 real datasets for fair evaluation.
- TGAN outperforms Bayesian network baselines on most real datasets, while other deep learning methods fail to achieve similar performance.
- The proposed method presents an innovative approach for modeling tabular data using conditional GANs.
- It addresses challenges posed by mixed discrete-continuous columns and imbalanced distributions.
- TGAN demonstrates superior performance compared to existing statistical and deep learning models on real-world datasets.

## Summary

The paper titled "Modeling Tabular data using Conditional GAN" addresses the challenging task of modeling the probability distribution of rows in tabular data and generating realistic synthetic data. Tabular data typically consists of a combination of discrete and continuous columns, with continuous columns potentially having multiple modes and discrete columns being imbalanced. Existing statistical and deep neural network models struggle to accurately model this type of data. To overcome these challenges, the authors propose TGAN, a method that utilizes a conditional generative adversarial network. TGAN is designed to effectively model the complex distribution of tabular data by capturing both the discrete-continuous nature and the potential imbalance in the columns. To ensure a fair and comprehensive evaluation, the authors create a benchmark consisting of 7 simulated datasets and 8 real datasets. They compare TGAN against several Bayesian network baselines. The results show that TGAN outperforms Bayesian methods on most of the real datasets, while other deep learning methods fail to achieve similar performance. Overall, this paper presents an innovative approach for modeling tabular data using conditional GANs, addressing the challenges posed by mixed discrete-continuous columns and imbalanced distributions. The proposed method demonstrates superior performance compared to existing statistical and deep learning models on real-world datasets.