A Survey on Tabular Data Generation: Utility, Alignment, Fidelity, Privacy, and Beyond

TL;DR

The paper surveys deep generative methods for tabular data through four user-centric requirements: $\text{utility}$, $\text{alignment}$, $\text{fidelity}$, and $\text{privacy}$. It formalizes tabular data as $D=(\mathcal{X},\mathcal{A},\mathcal{V})$ and, by assuming the real data distribution $p_X$, discusses learning a model distribution $p_\theta$ that approximates $p_X$, with $p_\theta$ increasingly used across GAN-, diffusion-, flow-, VAE-, and LLM-based architectures. The survey provides a taxonomy of methods and detailed evaluation protocols and metrics for each requirement, including $TSTR$ for utility, constraint-based metrics for alignment, feature/pair/joint fidelity metrics, and privacy attack-based metrics. A key finding is that no existing model simultaneously satisfies all four requirements, highlighting the need for holistic, alignment-aware evaluation and the development of hybrid models that balance utility, fidelity, alignment, and privacy in real-world deployments.

Abstract

Generative modelling has become the standard approach for synthesising tabular data. However, different use cases demand synthetic data to comply with different requirements to be useful in practice. In this survey, we review deep generative modelling approaches for tabular data from the perspective of four types of requirements: utility of the synthetic data, alignment of the synthetic data with domain-specific knowledge, statistical fidelity of the synthetic data distribution compared to the real data distribution, and privacy-preserving capabilities. We group the approaches along two levels of granularity: (i) based on the primary type of requirements they address and (ii) according to the underlying model they utilise. Additionally, we summarise the appropriate evaluation methods for each requirement and the specific characteristics of each model type. Finally, we discuss future directions for the field, along with opportunities to improve the current evaluation methods. Overall, this survey can be seen as a user guide to tabular data generation: helping readers navigate available models and evaluation methods to find those best suited to their needs.

Figures (2)

• Figure 1: Visualisation of common model types used for synthesising tabular data. Here, $x$, $z$, and $\tilde{x}$ denote a real sample, noise sample, and synthetic sample, respectively; $\tilde{x}'$ is a sample modified by layer-based models; $x_i$ is the sample at step $i$ of forward diffusion (for $i$ in $\{1,\ldots, T\}$, where $T$ is the maximum number of diffusion steps); and $\tilde{x}_i$ is the sample at step $i$ of reverse diffusion.
• Figure 2: Visualisation of the surveyed models based on the requirements they address.