Research map / Image generation / Fast diffusion sampling
TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TabDiff, a joint continuous-time mixed-type diffusion model (Gaussian VE diffusion for numerical columns, masked diffusion for categorical columns) with per-feature learnable noise schedules, a stochastic sampler, and classifier-free guidance for imputation, evaluated on seven UCI tabular datasets.
Problem
Generating high-quality synthetic mixed-type tabular data is hard due to heterogeneous feature types, complex inter-column correlations, and differing column-wise marginal distributions.
Approach
A joint continuous-time diffusion over native numerical and categorical columns (VE Gaussian plus absorbing-mask diffusion) with feature-wise learnable noise schedules trained end-to-end, a mixed-type stochastic sampler that re-noises to correct decoding errors, and a classifier-free guidance formulation for mixed-type conditional imputation.
Key results
- Average Shape error rate 1.17% vs 1.35% for TabSyn (about 13.3% relative improvement), and average Trend error 1.80% vs 2.33% for TabSyn (22.6% improvement).
- Average alpha-Precision of 98.22 (ranked first) across seven datasets, versus 97.86 for TabSyn.
- Average MLE gap 5.76% vs 6.78% for TabSyn, and imputation with CFG (omega=0.6) gives 5.60% average improvement over XGBoost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TabDiff instead of TabSyn | Tabular data generation · Seven datasets (average Shape error) | Shape error: 1.35 -> 1.17; 13.3% ↓ | Helped |
| Use TabDiff instead of TabSyn | Tabular data generation · Seven datasets (average Trend error) | Trend error: 2.33 -> 1.80; 22.6% ↓ | Helped |
| Use TabDiff instead of TabSyn | Synthetic data for downstream tasks · Seven datasets (average gap) | MLE average gap: 6.78 -> 5.76 | Helped |
| Use TabDiff instead of TabSyn | Tabular data generation · Seven datasets (average) | α-Precision: 97.86 -> 98.22 | Helped |
| Use TabDiff instead of TabSyn | Tabular data generation · Seven datasets (average) | β-Recall: 49.40 | Helped |
| Use classifier-free guidance with ω=0.6 instead of ω=0.0 (ablation) | TabDiff missing value imputation · Seven datasets (average) | Improvement over XGBoost: 3.76 -> 5.60 | Helped |
| Add stochastic sampling instead of deterministic ODE sampling (ablation) | TabDiff · Shape and Trend fidelity metrics | Shape and Trend | Helped |
| Add learnable feature-wise noise schedules instead of fixed schedules (ablation) | TabDiff · Tabular data generation | Trend | Helped |
| Add learnable feature-wise noise schedules instead of fixed schedules (ablation) | TabDiff on Adult · Training | Numerical and categorical training losses | Helped |
Surprising findings
- Discrete-time baselines like TabDDPM produce degenerate outputs on News and Diabetes (e.g., alpha-Precision 0.00 on News), per the authors' reported results.
- MLE scores can be very close across methods with differing fidelity scores, so the authors argue MLE is not a reliable indicator of data quality.
Limitations
- MLE scores may not reliably reflect data quality, as acknowledged by the authors.
- Some baseline results were reproduced by the authors or taken from prior work, and several baselines hit OOM on larger datasets.
- Evaluation limited to seven UCI datasets.
Experimental setup
Datasets: Adult Default Shoppers Magic Beijing News Diabetes
Benchmarks: Adult Default Shoppers Magic Beijing News Diabetes
Metrics: Shape Trend alpha-Precision beta-Recall Detection (C2ST) machine learning efficiency (AUC/RMSE) missing value imputation accuracy DCR
Baselines: CTGAN TVAE GOGGLE GReaT STaSy CoDi TabDDPM TabSyn
Builds on: TabSyn TabDDPM EDM D3PM
Closest papers by meaning
- Continuous Diffusion for Mixed-Type Tabular Data (ICLR 2025 · similarity 0.92)
- TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data (ICML 2025 · similarity 0.91)
- Mixed-Type Tabular Data Synthesis with Score-based Diffusion in Latent Space (ICLR 2024 · similarity 0.90)
- Understanding and Mitigating Memorization in Diffusion Models for Tabular Data (ICML 2025 · similarity 0.90)
- ClavaDDPM: Multi-relational Data Synthesis with Cluster-guided Diffusion Models (NeurIPS 2024 · similarity 0.89)
- CTSyn: A Foundation Model for Cross Tabular Data Generation (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.