PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey of Synthetic Tabular Data Generation

Ruxue Shi, Yili Wang, Mengnan Du, Xu Shen, Yi Chang, Xin Wang

arXiv:2504.16506v3cs.LG

TL;DR

Synthetic tabular data use is constrained by privacy restrictions, missing values, and class imbalance. This survey reviews the generation pipeline and compares traditional, diffusion, and LLM-based methods; its experiments find higher data availability for diffusion and LLM approaches than traditional methods, with LLM hallucinations challenging pairwise fidelity.

  • Problem

    Practical use of tabular data is constrained by privacy restrictions, missing values, and imbalanced class distributions.

    Source: §I · p. 1

  • Method

    The survey reviews the end-to-end generation process, categorizes traditional, diffusion, and LLM-based methods, and synthesizes post-processing and evaluation approaches.

    Source: §I · pp. 1–2

  • Results

    Diffusion and LLM-based methods show markedly higher data availability than traditional methods, but LLM hallucinations challenge pairwise fidelity.

    Source: D. Analysis and Discussion · p. 12

  • Takeaways & Limitations

    The survey aims to provide practical guidance for organizations and summarize evaluation protocols, real-world applications, and future research directions.

    Source: §VII · p. 14

Abstract

from arXiv · show

Tabular data is one of the most prevalent and important data formats in real-world applications such as healthcare, finance, and education. However, its effective use in machine learning is often constrained by data scarcity, privacy concerns, and class imbalance. Synthetic tabular data generation has emerged as a powerful solution, leveraging generative models to learn underlying data distributions and produce realistic, privacy-preserving samples. Although this area has seen growing attention, most existing surveys focus narrowly on specific methods (e.g., GANs or privacy-enhancing techniques), lacking a unified and comprehensive view that integrates recent advances such as diffusion models and large language models (LLMs). In this survey, we present a structured and in-depth review of synthetic tabular data generation methods. Specifically, the survey is organized into three core components: (1) Background, which covers the overall generation pipeline, including problem definitions, synthetic tabular data generation methods, post processing, and evaluation; (2) Generation Methods, where we categorize existing approaches into traditional generation methods, diffusion model methods, and LLM-based methods, and compare them in terms of architecture, generation quality, and applicability; and (3) Applications and Challenges, which summarizes practical use cases, highlights common datasets, and discusses open challenges such as heterogeneity, data fidelity, and privacy protection. This survey aims to provide researchers and practitioners with a holistic understanding of the field and to highlight key directions for future work in synthetic tabular data generation.

I. INTRODUCTION

The survey addresses gaps in prior reviews by integrating synthetic tabular generation methods, post-processing, evaluation, applications, and future directions. It organizes methods into traditional, diffusion, and LLM-based approaches and frames evaluation around data availability and privacy.

I. INTRODUCTION

  • Existing surveys cover narrow domains or methods, often omit post-processing, and have not thoroughly examined diffusion and LLM advances.Synthetic data can match a source distribution yet violate commonsense or logical rules, such as assigning negative ages or marking children as married.Source: §I · p. 1
  • The survey categorizes traditional, diffusion, and LLM-based methods and compares their strengths and limitations.Its comparison spans data-type compatibility, sample fidelity, privacy resilience, and downstream performance.Source: §I · pp. 1–2
  • The review also covers real-world applications and outlines promising directions for future research.It aims to provide practical guidance for organizations considering synthetic tabular data generation.Source: §I · pp. 1–2
  • It introduces a post-processing taxonomy that distinguishes sample enhancement from label enhancement.The survey presents these strategies as addressing synthetic-data quality and alignment with human knowledge.Source: §I · pp. 1–2
  • It organizes evaluation around data availability and privacy protection, reflecting synthetic data’s dual objectives.Availability examples include machine-learning efficiency, fidelity, and alignment; privacy examples include membership inference and DCR.Source: §I · pp. 1–2
  • The survey compiles representative papers, open-source implementations, datasets, and evaluation tools in a curated repository.The repository is intended to promote reproducibility.Source: §I · p. 2

II. DEFINITION

  • The paper introduces the generation pipeline and identifies its task definition and associated challenges as background for the review.The pipeline trains a generator, post-processes samples, and evaluates data availability and privacy.Source: §II · p. 2, Figure 1 · p. 2

A. Problem Definition

Synthetic tabular generation learns a dataset’s statistical distribution and produces a structurally matching synthetic dataset. Its practical challenges include limited, imbalanced, and incomplete data, while post-processing and evaluation address validity, availability, and privacy.

A. Problem Definition

  • The task is to learn the statistical distribution of a real tabular dataset and produce a synthetic dataset that closely approximates it.The real and synthetic datasets share a feature-name structure; rows contain feature values and corresponding labels.Source: A. Problem Definition · p. 2
  • Post-processing addresses mislabeled samples and violations of human knowledge or domain constraints, while evaluation covers data availability and privacy protection.Availability metrics include machine-learning efficiency, fidelity, and alignment; privacy metrics include DCR, attribute inference, and membership inference attacks.Source: A. Problem Definition · p. 2

B. Challenges for Synthetic Tabular Data Generative Models

  • Heterogeneous, structurally diverse tabular data challenges effective modeling and high-quality synthesis.The section frames its challenges around data quality, inherent characteristics, and complex distributions.Source: B. Challenges for Synthetic Tabular… · p. 3

1) Data Quality:

  • Limited dataset size, class imbalance, and missing values complicate training and can impair minority-class representation.Scarcity restricts data availability and model training; imbalanced data can yield biased synthetic outputs, while incomplete observations increase modeling complexity.Source: 1) Data Quality · p. 3

2) Inherent Characteristics:

Tabular data poses challenges through mixed feature types, sparse inter-column relationships, and complex distributions. Numerical and categorical imbalances, column-dependent meaning, and non-Gaussian values make faithful synthesis difficult.

2) Inherent Characteristics:

  • Mixed numerical and categorical columns have distinct statistical properties, complicating joint-distribution modeling.Some columns may themselves mix numerical and categorical values, challenging encoding and coherent representation.Source: 2) Inherent Characteristics · p. 3
  • Sparse inter-column dependencies include both same-type and mixed-type relationships, so preserving realism requires capturing which column pairs are related.For example, age and education may be interrelated while other pairs remain independent.Source: 2) Inherent Characteristics · p. 3

3) Complex data Distributions:

  • Skewed, heavy-tailed, or multimodal numerical values challenge models that assume Gaussian distributions.The passage identifies non-Gaussian numerical columns as a challenge for learning complex distributions.Source: 3) Complex data Distributions · pp. 3–4
  • Imbalanced categorical columns can lead models to overfit dominant categories and underrepresent minority categories, reducing synthetic-data diversity.The meaning of a value can also depend on its column: “10” may represent age or a rating.Source: 3) Complex data Distributions · p. 4

III. SYNTHETIC TABULAR DATA GENERATION METHODS

The survey organizes generation methods into traditional, diffusion, and LLM-based families, then reviews representative approaches and their strengths and limitations.

III. SYNTHETIC TABULAR DATA GENERATION METHODS

  • The survey’s taxonomy separates traditional generation, diffusion models, and LLM-based methods.It introduces representative approaches in each category and analyzes their strengths and limitations.Source: §III · p. 4

1) Machine Learning-based Methods:

2) VAE-based Method:

  • VAEs improve synthetic tabular data quality by learning latent representations while maximizing the Evidence Lower Bound.The encoder and decoder are represented by q(z|x) and p(x|z), respectively.Source: 2) VAE-based Method · p. 5
  • TVAE encodes categorical features with one-hot vectors, VAEM uses two training stages, and GOGGLE learns inter-column dependencies with a graph adjacency matrix.VAE-generated data can tend toward averaging and lose detail.Source: 2) VAE-based Method · p. 5

3) GAN-based Generators:

The survey describes GAN-based tabular generators as traditional, conditional, or differentially private, with extensions addressing data types, feature constraints, and privacy.

Traditional GAN

  • Traditional GANs adapt architectures to tabular data, including LSTM-based feature generation, autoencoder-based EHR synthesis, and convolutional generators.Other designs use sequential generators, feature tokenization, Fourier networks, or fairness constraints; these methods lack generation conditioned on specific feature constraints.Source: 3) GAN-based Generators · p. 5

Conditional GAN

  • Conditional GANs add conditioning information to control generated features, supporting class-specific synthesis and minority-class oversampling.CTGAN uses training-by-sampling for imbalanced categorical features, while RC-TGAN conditions child-table generation on parent rows.Source: 3) GAN-based Generators · pp. 5–6

Differentially Private GAN

  • Differentially private GAN training clips per-sample gradients, averages them, and adds calibrated Gaussian noise to protect privacy while trading off model utility.The survey also describes PATE-GAN, RDP-CGAN, and CTAB-GAN+ as privacy-oriented GAN approaches.Source: 3) GAN-based Generators · p. 6

B. Diffusion Model Methods

Diffusion models use denoising to generate tabular data and are presented as promising for complex distributions, with adaptations for heterogeneous features and domain-specific data.

B. Diffusion Model Methods

  • Diffusion models’ denoising mechanism addresses GAN challenges such as training instability, mode collapse, and poor multimodal representation.The survey reports notable performance improvements and potential for diverse, high-quality synthetic data.Source: B. Diffusion Model Methods · p. 6

1) DDPM-based Method:

  • DDPMs learn generation through forward noising and reverse denoising, with model parameters optimized using a variational lower bound on data log-likelihood.The forward process iteratively adds noise; the reverse process learns transitions that recover data from noise.Source: 1) DDPM-based Method · pp. 6–7
  • Tabular diffusion variants address heterogeneous features with joint continuous-time processes, separate models for numerical and categorical data, or unified latent representations.Other approaches use feature-wise learnable processes, autoencoders, or lightweight controllers for conditional generation.Source: 1) DDPM-based Method · p. 7
  • Diffusion applications include EHR synthesis and financial data generation, including approaches targeting missing values and minority categories.Examples include TabDDPM-EHR, FLEXGEN-EHR, FinDiff, and Imb-FinDiff.Source: 1) DDPM-based Method · p. 7

2) Score-based Method:

Score-based diffusion methods reverse a continuous-time noising process using an estimated score function. The survey also situates LLM-based generation as a response to models’ limited use of domain knowledge and logical constraints.

2) Score-based Method:

  • Score-based models define a continuous-time forward diffusion and reverse it using the gradient of the log data density.The score function is estimated with a neural network because the density is unknown.Source: 2) Score-based Method · p. 8
  • Training uses denoising score matching, and generation starts from Gaussian noise before integrating the reverse SDE backward in time.The survey names Euler-Maruyama and Predictor-Corrector schemes as numerical solvers.Source: 2) Score-based Method · p. 8
  • Tabular score-based methods address varied needs, including missing values, mixed feature types, sampling stability, and latent-space generation.Examples include MissDiff for incomplete data, CSDI-T for categorical and numerical features, STaSy for training stability, and TABSYN for VAE latent-space diffusion.Source: 2) Score-based Method · p. 8

C. LLM-based Methods

  • LLM-based tabular generation is explored because raw-distribution learning may miss domain knowledge and produce samples that violate logical constraints.Methods are grouped by whether the LLM is fine-tuned.Source: C. LLM-based Methods · p. 8

1) Prompt-based Methods:

The survey distinguishes prompting from fine-tuning for LLM-based synthetic tabular generation, then describes post-processing approaches for improving sample quality and correcting labels.

1) Prompt-based Methods:

  • Prompt-based methods serialize tabular data for LLM inputs and use designed prompts to generate synthetic datasets without fine-tuning.EPIC guides generation with balanced grouped examples; CLLM curates augmented data, while LITO targets minority-class oversampling.Source: 1) Prompt-based Methods · p. 8, Figure 5 · p. 9

2) Fine-Tuning Methods:

  • Fine-tuning is used to reduce inaccurate or misleading outputs from direct prompting and improve the fit of generated data to dataset structure and constraints.GReaT fine-tunes an LLM for realistic tabular sampling and can adjust any subset of features.Source: 2) Fine-Tuning Methods · pp. 8–9
  • TabMT addresses heterogeneous fields and missing values, while TAPTAP uses tabular pre-training for applications including imputation and imbalanced classification.HARMONIC uses k-nearest-neighbor-inspired instruction data to teach inter-row relationships while targeting data format and connections rather than the data itself.Source: 2) Fine-Tuning Methods · p. 9, Table III · p. 9
  • AIGT uses metadata prompts and long-token partitioning for tabular data of any scale, while DP-LLMTGen uses two-stage fine-tuning for differentially private synthesis.P-TA applies PPO to guide LLM feature distributions toward more realistic data.Source: 2) Fine-Tuning Methods · p. 9

IV. POST PROCESSING

  • Synthetic samples may be unrealistic, incorrect, or toxic, so post-processing methods are grouped into sample enhancement and label enhancement.The stated causes include complex tabular distributions, limited common-sense knowledge in non-LLM methods, and LLM hallucinations.Source: §IV · pp. 9–10, Figure 6 · p. 10

A. Sample Enhancement

  • Sample enhancement modifies feature values or filters unreasonable samples, with filtering and correction as its two approaches.Filtering criteria can use learning dynamics, confidence, uncertainty, or self-authentication with rejection sampling.Source: A. Sample Enhancement · p. 10, 1) Sample Filtering · p. 10

1) Sample Filtering:

  • Sample filtering removes samples flagged as uncertain or inconsistent by confidence-based or model-based checks.CLLM uses confidence and uncertainty, while LITO uses self-authentication and rejection sampling.Source: 1) Sample Filtering · p. 10

2) Sample Correction:

The survey covers correction of synthetic samples and labels, then outlines evaluation dimensions, common datasets, and a downstream-learning protocol for measuring synthetic-data utility.

2) Sample Correction:

  • Linear-inequality constraints can align synthetic samples with background knowledge, while QFLRA extends constraint representation to non-convex and disconnected spaces.The linear-inequality approach is limited to knowledge expressible in that form.Source: 2) Sample Correction · p. 10

B. Label Enhancement

  • Label enhancement uses manual re-labeling or proxy models to correct potential annotation errors in synthetic samples.Manual review can be costly and impractical at scale; TAPTAP and AIGT use backbone classifiers for pseudo-labels, while Pred-LLM prompts an LLM for labels.Source: B. Label Enhancement · p. 10

V. EVALUATION

  • Evaluation considers data availability and privacy protection, including downstream-task support, distribution fidelity, domain-knowledge alignment, and disclosure risk.These are presented as complementary evaluation perspectives.Source: §V · p. 11

A. Dataset

  • The survey summarizes 20 widely used datasets across regression, binary-classification, and multi-class-classification tasks, drawn mainly from UCI, OpenML, and Kaggle.Beijing, Adult, and California Housing are noted for studies involving missing values.Source: A. Dataset · p. 11, Table IV · p. 11

B. Data Availability

  • Data availability utility comprises ML efficiency, fidelity to original statistical properties, and alignment with human domain knowledge.These dimensions target data augmentation needs involving insufficient data, class imbalance, and missing values.Source: B. Data Availability · p. 11

1) ML Efficiency:

  • TSTR assesses ML utility by training separate models on synthetic and real training data and evaluating both on the same real test set.Comparable performance, using measures such as accuracy, F1, AUC, or RMSE, indicates high synthetic-data utility under this protocol.Source: 1) ML Efficiency · p. 11

0 ROC(t) d(t) is commonly used for evaluating categorical

The survey organizes fidelity evaluation around feature-wise, feature-pair, and joint-distribution comparisons, while noting that fidelity alone does not establish downstream utility.

2) Fidelity:

  • High fidelity does not always imply strong downstream task performance, although fidelity remains a fundamental measure of structural realism.The survey cites prior studies for this qualification.Source: 2) Fidelity · pp. 11–12
  • Column-wise fidelity compares individual features, using KST for numerical distributions and TVD for categorical frequencies.KST measures the maximum discrepancy between cumulative distribution functions; TVD compares category probabilities.Source: 2) Fidelity · p. 12
  • Pair-wise fidelity evaluates whether dependencies between feature pairs are preserved using measures suited to numerical, categorical, or mixed features.Examples include Pearson correlation matrices and differences between contingency similarity matrices.Source: 2) Fidelity · p. 12
  • Joint-distribution fidelity compares full real and synthetic distributions using measures such as β-Recall and Coverage score.The survey identifies these as joint-distribution metrics.Source: 2) Fidelity · p. 12

3) Alignment:

The survey describes alignment and privacy measures for evaluating synthetic data, then compares representative generation methods and frames practical applications around data availability and privacy protection.

3) Alignment:

  • CVR measures the share of samples violating constraints, CVC the share of constraints violated, and sCVC the average per-constraint share of violating samples.These are presented as measures of alignment with background knowledge.Source: 3) Alignment · p. 12

C. Privacy Protection

  • Privacy evaluation covers record similarity, attribute inference, and membership inference to assess resemblance, sensitive-attribute leakage, and training-set membership risks.A small DCR can indicate excessive similarity to real individuals; high attribute- or membership-inference success signals privacy risk.Source: C. Privacy Protection · p. 12

D. Analysis and Discussion

  • Diffusion and LLM methods yield markedly higher data availability than traditional methods, while LLM hallucinations challenge pairwise fidelity.Diffusion performs better with fewer columns; LLM-based methods handle more columns, while diffusion’s advantage diminishes as categorical features increase.Source: D. Analysis and Discussion · pp. 12–13, Table V · p. 13

VI. PRACTICAL APPLICATIONS AND FUTURE DIRECTIONS

  • The survey frames practical applications from two complementary perspectives: improving tabular-data availability and privacy-protective data analysis.It also presents representative applications and future research directions.Source: §VI · p. 13, A. Practical Applications · p. 13

1) Improving Tabular Data Availability:

The survey describes synthetic data approaches for addressing imbalance, missing values, and privacy-sensitive use, and identifies unresolved challenges in balancing utility and privacy, human alignment, interpretability, and evaluation.

1) Improving Tabular Data Availability:

  • Conditional generation and score-based diffusion target minority classes, while imputation methods, diffusion models, and LLMs address missing values.MICE estimates conditional feature distributions iteratively; diffusion methods condition on missing-column names, and fine-tuned LLMs can infer missing values through prompts.Source: 1) Improving Tabular Data Availability · p. 13

2) Privacy Protection:

  • Synthetic EHR studies report limited identity and attribute-disclosure risks, while diffusion methods target realistic records and FLEXGEN-EHR addresses missing modalities.The cited work evaluates generated EHRs through distribution comparisons, predictive tasks, and medical-professional assessments.Source: 2) Privacy Protection · pp. 13–14

B. Future Directions

  • Early methods often prioritize either data utility or privacy, leaving specialized models with limited generalizability and a challenge of balancing both objectives.Source: B. Future Directions · p. 14
  • Matching the original data distribution does not guarantee consistency with human commonsense or domain-specific logic.The survey identifies human-knowledge alignment as a research direction because misalignment can pose risks in real-world use.Source: B. Future Directions · p. 14
  • Black-box generative models make internal decisions difficult to understand and may inherit or amplify source-data biases.The survey calls for interpretability methods to support transparency, fairness, and trustworthiness, particularly in high-stakes applications.Source: B. Future Directions · p. 14
  • Synthetic tabular data quality is harder to assess than image quality and often requires domain-specific validation against real-world constraints and expert knowledge.Source: B. Future Directions · p. 14

VII. CONCLUSION

  • The survey aims to guide organizational adoption while summarizing evaluation protocols, real-world applications, and future research directions.Source: §VII · p. 14
Loading 2504.16506v3…