Research map / Probabilistic and scientific modeling / Molecules and proteins

TabSDS: a Lightweight, Fully Non-Parametric, and Model Free Approach for Generating Synthetic Tabular Data

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Prediction-powered and conformal inference

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 TabSDS, a lightweight, non-parametric, model-free synthetic tabular data generator that combines interpolated order-statistic marginal sampling with SJPPDS restricted-permutation shuffling and rank matching, and benchmarks it against Synthcity generators.

Problem

Deep generative models for synthetic tabular data are computationally heavy and require extensive hyperparameter tuning, while lightweight alternatives are limited; the paper targets fast, tunable synthetic tabular data generation that preserves joint distribution and privacy trade-offs.

Approach

Extends the SJPPDS data-shuffling perturbation method to categorical and mixed data and to full synthesis: it generates uncorrelated synthetic marginals via interpolated order statistics (IOSSampler), then rank-matches them to restricted-permutation-shuffled real data so the association structure is recovered, with the number of levels nc as the single key tuning knob for the fidelity-privacy trad

Key results

• TabSDS ranked first or second on most datasets for fidelity and utility against Synthcity baselines including DDPM and ARF.
• TabSDS runtime is orders of magnitude faster than DDPM, TVAE and CTGAN and faster than ARF in most datasets (e.g., 22.98 s vs 116.92 s DDPM on Diabetes 130US).
• On the Mushroom dataset, TabSDS achieved detection AUROC 0.664 versus 0.954 for DDPM, and on Abalone TabSDS achieved 0.837 versus DDPM 0.878.

Measured effects reported in the paper

Surprising findings

• On the Electricity dataset TabSDS had a much worse detection-test AUROC (0.976) than DDPM (0.802), and on Credit ML efficiency was lower than DDPM, showing it is not uniformly best.

Limitations

• Model-free design cannot be adapted to fair synthetic data generation (unlike DECAF).
• IOSSampler may generate a small fraction of values outside the training range for discontinuous multimodal distributions.
• Only two tuning parameters, but nc must be chosen by the analyst via trade-off inspection; selection criteria were dataset-specific.

Experimental setup

Datasets: Abalone Adult Bank marketing California housing Credit Diabetes 130US Electricity Eye movements House 16H Magic telescope Mushroom Pol

Benchmarks: Synthcity benchmarks (fidelity, utility, privacy) across 12 tabular datasets DCR, SDBRL, SSDID privacy metrics

Metrics: detection test AUROC (XGBoost) ML efficiency (AUROC / R2) Domias MIA AUROC distance to closest record runtime in seconds

Baselines: TabDDPM ARF TVAE CTGAN BayesNet SMOTE

Builds on: SJPPDS rank matching (Domingo-Ferrer and Muralidhar 2016) CART synthesis

Closest papers by meaning

• How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.89)
• TabEBM: A Tabular Data Augmentation Method with Distinct Class-Specific Energy-Based Models (NeurIPS 2024 · similarity 0.88)
• TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data (ICML 2025 · similarity 0.88)
• TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.87)
• TabM: Advancing tabular deep learning with parameter-efficient ensembling (ICLR 2025 · similarity 0.87)
• CTSyn: A Foundation Model for Cross Tabular Data Generation (ICLR 2025 · similarity 0.87)