🤖 machine learning

TDGT: A Tabular Data Generation Toolkit supporting adaptive GPU-accelerated Bayesian mixture models, diffusion-based models, and latent-space generative modeling

This paper introduces TDGT, a web-based toolkit for synthetic tabular data generation that features an adaptive, GPU-accelerated Bayesian mixture model (ABMS) and a hybrid VAE-ABMS architecture to automate hyperparameter tuning and ensure high-fidelity, privacy-preserving data synthesis across diverse domains.

Original authors: Vasileios C. Pezoulas, Nikolaos S. Tachos, Eleni Georga, Kostas Marias, Manolis Tsiknakis, Dimitrios I. Fotiadis

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vasileios C. Pezoulas, Nikolaos S. Tachos, Eleni Georga, Kostas Marias, Manolis Tsiknakis, Dimitrios I. Fotiadis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor, a bank manager, or a cybersecurity expert. You have a treasure trove of data—patient records, transaction logs, or network traffic—but you can't share it with the world because it contains private secrets. You need to share the patterns and lessons from the data without revealing the people behind them.

This is where TDGT comes in. Think of TDGT as a "Digital Photocopier for Data" that doesn't just copy the pages; it writes an entirely new book that looks, feels, and acts exactly like the original, but every single sentence is made up.

Here is how the paper explains this toolkit, broken down into simple concepts:

1. The Problem: The "Manual" Bottleneck

Until now, making this kind of fake data was like trying to bake a perfect cake without a recipe. You needed to be a master chef (a computer scientist) to tweak the ingredients (hyperparameters) manually. If you got the temperature wrong, the cake (the data) would be ruined. Also, there was no built-in taste-tester to tell you if the fake cake actually tasted like the real one.

TDGT changes this. It's a web-based toolkit (like a website you can click on) that does the baking for you. You just upload your data, pick a "baking style," and it spits out a perfect fake dataset with a report card telling you how good it is. No coding required.

2. The Secret Sauce: Three Ways to Bake

The paper introduces a "menu" of different methods to create this fake data, ranging from simple to complex.

The "Smart Cluster" Baker (ABMS):
Imagine you have a bag of mixed jellybeans. A simple baker might just say, "It's all red and blue." But the Adaptive Bayesian Mixture Synthesizer (ABMS) is a smart baker. It looks at the bag and automatically figures out, "Ah, there are actually 5 distinct flavors here, not just 2." It counts the clusters for you automatically, so you don't have to guess. It's fast and great for simple data.
The "Deep Dive" Baker (VAE-ABMS):
Some data is like a complex puzzle where the pieces are twisted and tangled in weird ways. The VAE-ABMS takes the data, flattens it out into a simpler "shadow world" (latent space), counts the clusters there, and then builds the fake data back up. This is great for data with complicated, non-linear relationships.
The "Super-Speed" Baker (ABMS-CUDA):
If you have a massive bag of jellybeans (millions of rows), the smart baker might get tired. ABMS-CUDA is the same baker, but with a supercomputer (GPU) attached to their brain. It does the counting 3 to 4 times faster, making it perfect for huge datasets.

3. The "Deep Learning" Chefs

For the most complex data, TDGT also includes three advanced "Deep Learning" chefs:

TabGAN: A chef that plays a game of "fake it 'til you make it" against a critic, constantly improving until the fake data is indistinguishable from the real thing.
TabVAE: A chef that learns to compress the data into a summary and then expand it back out, creating new variations.
TabDiffusion: A chef that starts with pure noise (static) and slowly "denoises" it, step-by-step, until a clear picture of the data emerges.

4. The Taste Test: Did it Work?

You can't just trust the baker; you need to taste the cake. TDGT has a built-in Quality Control Lab that runs 11 different tests:

Distribution Checks: Does the fake data have the same shape as the real data? (e.g., if real data has a few very high values, does the fake data too?)
Relationship Checks: If "Age" and "Salary" go up together in the real data, do they still go up together in the fake data?
Privacy Checks: Does the fake data accidentally contain a real person's exact record? (It checks for things like "k-anonymity" to ensure no one can be re-identified).

5. The Results: What Worked Best?

The authors tested their toolkit on three real-world scenarios:

Healthcare: Breast cancer records (small but complex).
Finance: Bank marketing data (medium size, mixed types).
Cybersecurity: Network attack logs (huge size, very messy).

The Findings:

Speed vs. Quality: If you need data right now (in seconds), the ABMS (Smart Cluster) method is the winner. It's incredibly fast and "good enough" for many tasks.
High Fidelity: If you need the data to be perfectly accurate for complex research, the TabDiffusion and VAE-ABMS methods are the best. They capture the subtle, twisted relationships in the data better than the others, though they take longer (minutes instead of seconds) to bake.
The GPU Boost: For huge datasets, the ABMS-CUDA (Super-Speed) version was a game-changer, cutting the time down from over a minute to just a few seconds without losing quality.

6. The Bottom Line

TDGT is a one-stop shop. It takes the complicated math of creating fake data, hides it behind a simple website, and gives you a result that is statistically sound and privacy-safe. It bridges the gap between "experts who know how to code" and "practitioners who just need the data," allowing anyone to generate high-quality synthetic data for research and analysis without needing a PhD in computer science.

Technical Summary: TDGT – A Tabular Data Generation Toolkit

Problem Statement

The proliferation of AI-driven systems in regulated domains (healthcare, finance, cybersecurity) has intensified the demand for high-quality synthetic tabular data to facilitate privacy-preserving data sharing. While synthetic data offers a robust alternative to traditional anonymization by generating entirely new samples rather than perturbing existing ones, current tools present significant barriers to adoption for domain researchers. Existing solutions often require:

Programming expertise and scripting fluency.
Manual hyperparameter configuration (e.g., selecting the number of mixture components or latent dimensions), which demands deep generative modeling knowledge.
Disparate evaluation pipelines, lacking integrated, multi-metric assessment of statistical fidelity and privacy risk.

Consequently, there is a gap between methodological advances in generative modeling and the practical accessibility of these tools for practitioners who need synthetic data but lack specialized modeling expertise.

Methodology

The paper introduces TDGT (Tabular Data Generation Toolkit), a web-based, end-to-end platform that unifies data generation, evaluation, and visualization. The toolkit implements a four-stage modular pipeline:

Data Ingestion & Feature Detection: Automated classification of columns as categorical, integer, or continuous, with metadata extraction (category registries, precision, bounds).
Synthetic Data Generation: TDGT supports nine generators, including three parametric baselines (MVND, LogMVND, BGMM) and six advanced methods. The core methodological innovations are:
- Adaptive Bayesian Mixture Synthesizer (ABMS): A novel algorithm that autonomously determines the optimal number of Gaussian mixture components ( $k^*$ ) by iterating through candidate cluster counts and optimizing a cluster quality criterion (Silhouette score or Davies-Bouldin index). It employs an adaptive Dirichlet prior to suppress spurious components, eliminating manual hyperparameter tuning.
- VAE-ABMS: A hybrid architecture where a Variational Autoencoder (VAE) learns a compact latent representation of the input data. ABMS is then applied within this latent space to synthesize vectors, which are decoded back to the original feature space. This addresses non-linear dependencies that shallow mixture models cannot capture.
- ABMS-CUDA: A GPU-accelerated variant of ABMS leveraging CUDA-based k-means clustering and Gaussian mixture fitting for large-scale datasets. It includes a fallback mechanism to CPU execution if GPU resources are insufficient.
- Deep Generative Models: The toolkit integrates customized implementations of TabGAN (Wasserstein GAN with gradient penalty), TabVAE (VAE with adaptive batch sizing), and TabDiffusion (Denoising Diffusion Probabilistic Model with sinusoidal time embeddings).
Post-Generation Constraint Enforcement: A uniform procedure ensures structural validity by remapping categorical values to observed registries, casting integers, rounding floats to original precision, and clipping continuous values to empirical bounds.
Multi-Metric Evaluation: The toolkit assesses fidelity using seven statistical metrics (KS, KL, JS, TVD, Hellinger distance, Wasserstein distance, and correlation difference $\Delta C$ ) and two privacy indicators (k-anonymity and disclosure risk). Results are visualized via interactive Plotly charts.

Key Contributions

The paper makes four primary contributions:

ABMS Algorithm: A data-driven approach to Bayesian mixture synthesis that automatically calibrates the number of components, removing the need for manual hyperparameter specification.
VAE-ABMS Architecture: A hybrid model combining VAE representation learning with adaptive mixture synthesis to handle complex, non-linear tabular distributions.
ABMS-CUDA: A scalable, GPU-accelerated implementation of ABMS that significantly reduces training time for large datasets while maintaining algorithmic semantics.
TDGT Toolkit: A unified, web-based interface that integrates these generation methods with a comprehensive evaluation suite and real-time visualizations, requiring no programming expertise from the user.

Experimental Results

The toolkit was benchmarked across three datasets: Wisconsin Breast Cancer (small, high-dimensional, continuous), Bank Marketing (moderate scale, mixed types), and NSL-KDD (large scale, heterogeneous, imbalanced).

Fidelity: Deep generative models generally outperformed mixture-based approaches in preserving structural correlations and marginal distributions, particularly on complex datasets.
- TabDiffusion achieved the best structural fidelity (lowest $\Delta C$ ) on Bank Marketing and NSL-KDD.
- VAE-ABMS demonstrated superior marginal fidelity (lowest KS and JS) on the high-dimensional NSL-KDD and Wisconsin datasets.
- TabGAN excelled in correlation preservation on the Wisconsin dataset.
- ABMS/ABMS-CUDA provided competitive distributional metrics (JS) but struggled with structural correlation preservation ( $\Delta C$ ) compared to deep models, particularly on high-dimensional data.
Performance & Scalability:
- ABMS was the fastest generator (seconds to minutes), making it suitable for rapid exploration.
- ABMS-CUDA achieved 3–4× speedups over CPU ABMS on datasets with $N > 10,000$ , though it incurred a slight fidelity penalty on the 42-feature NSL-KDD dataset due to necessary PCA projection for numerical stability.
- TabGAN incurred the highest training costs (up to ~20 minutes for Bank Marketing) due to its adversarial training loop.
- TabDiffusion offered a favorable trade-off, achieving high fidelity with training times significantly lower than TabGAN on large datasets.

Significance and Claims

The paper positions TDGT as a bridge between advanced generative modeling research and practical application in regulated domains. Its significance lies in:

Accessibility: By automating hyperparameter selection (via ABMS) and providing a no-code web interface, TDGT lowers the barrier to entry for domain researchers (e.g., clinicians, financial analysts) who lack deep learning expertise.
Integrated Evaluation: It addresses the reproducibility gap by offering a unified, multi-metric evaluation framework that assesses both statistical fidelity and privacy risk, eliminating the need for disparate assessment pipelines.
Scalability: The inclusion of GPU-accelerated variants (ABMS-CUDA) and efficient deep models (TabDiffusion) ensures the toolkit is viable for datasets ranging from small clinical cohorts to large-scale network traffic logs.
Regulatory Utility: The toolkit is designed to support privacy-preserving collaboration in sectors where data access is strictly constrained, enabling the creation of surrogate datasets that retain statistical properties without exposing identifiable records.

The authors conclude that while deep generative models offer superior fidelity for complex data, the adaptive mixture approaches (ABMS/VAE-ABMS) provide a critical balance of speed, interpretability, and performance, making them essential components of a comprehensive synthetic data toolkit.