[Paper Review] CTAB-GAN: Effective Table Data Synthesizing

Zilong Zhao, Aditya Kunar+3 more|arXiv (Cornell University)|Feb 16, 2021

Generative Adversarial Networks and Image Synthesis24 references63 citations

TL;DR

CTAB-GAN is a conditional GAN for tabular data that handles mixed data types, long-tail and imbalanced distributions, improving ML utility and statistical similarity over prior methods.

ABSTRACT

While data sharing is crucial for knowledge development, privacy concerns and strict regulation (e.g., European General Data Protection Regulation (GDPR)) unfortunately limit its full effectiveness. Synthetic tabular data emerges as an alternative to enable data sharing while fulfilling regulatory and privacy constraints. The state-of-the-art tabular data synthesizers draw methodologies from generative Adversarial Networks (GAN) and address two main data types in the industry, i.e., continuous and categorical. In this paper, we develop CTAB-GAN, a novel conditional table GAN architecture that can effectively model diverse data types, including a mix of continuous and categorical variables. Moreover, we address data imbalance and long-tail issues, i.e., certain variables have drastic frequency differences across large values. To achieve those aims, we first introduce the information loss and classification loss to the conditional GAN. Secondly, we design a novel conditional vector, which efficiently encodes the mixed data type and skewed distribution of data variable. We extensively evaluate CTAB-GAN with the state of the art GANs that generate synthetic tables, in terms of data similarity and analysis utility. The results on five datasets show that the synthetic data of CTAB-GAN remarkably resembles the real data for all three types of variables and results into higher accuracy for five machine learning algorithms, by up to 17%.

Motivation & Objective

• Motivate synthetic tabular data as a privacy-preserving alternative under GDPR.
• Develop a GAN-based generator capable of modeling mixed continuous/categorical variables and missing values.
• Address long-tail and highly imbalanced distributions with novel encoding and training strategies.
• Incorporate a classifier and information loss to boost semantic integrity and training stability.
• Evaluate against state-of-the-art tabular generators on multiple datasets for ML utility, statistical similarity, and privacy approximations.

Proposed method

• Introduce CTAB-GAN, a conditional table GAN with generator, discriminator, and auxiliary classifier.
• Use a Mixed-type Encoder to represent mixed categorical-continuous variables and missing values as concatenated mode-based encodings.
• Incorporate information loss to align first- and second-order statistics between real and synthetic data.
• Add a classification loss via an auxiliary classifier to enforce semantic consistency in generated records.
• Employ a log-frequency sampling of the conditional vector to mitigate mode collapse for imbalanced variables.
• Preprocess long-tailed continuous variables with a log transform to improve mode learning in variational Gaussian mixtures.

Experimental results

Research questions

• RQ1 Can CTAB-GAN accurately model mixed-type tabular data (continuous, categorical, and mixed) including missing values?
• RQ2 Does CTAB-GAN improve ML analysis utility using synthetic data compared to state-of-the-art tabular GANs?
• RQ3 How well does CTAB-GAN preserve statistical similarity (distributions and correlations) to real data?
• RQ4 What is the privacy risk associated with CTAB-GAN-generated data compared to peers?

Key findings

• CTAB-GAN outperforms CTGAN, TableGAN, CW-GAN and MedGAN in ML utility across five datasets (accuracy, F1, AUC differences).
• CTAB-GAN achieves better statistical similarity (lower JSD, lower WD, closer correlations) than competitors on average.
• CTAB-GAN shows stronger privacy related metrics (distance-based) than TableGAN and others, indicating reduced privacy risk while maintaining utility.
• Ablation studies confirm that classifier, information loss, and mixed-type encoding contribute to performance gains across datasets.