Research map / Federated learning and privacy / Differential privacy
CuTS: Customizable Tabular Synthetic 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.
Introduces CuTS, a framework for customizable tabular synthetic data generation that pre-trains a marginal-matching generator and fine-tunes it on differentiable relaxations of declarative privacy, logical, statistical, and downstream specifications.
Problem
Existing tabular synthetic data methods support only specific constraints (e.g., differential privacy or fairness) and cannot be freely customized with logical, statistical, or downstream requirements while keeping high utility.
Approach
Pre-train a GAN-style generator on three-way marginals (optionally with DP via iterative AIM-style training), then fine-tune it on a differentiable loss automatically derived from declarative specifications, using novel relaxations of row constraints, implications, conditional statistics, and a differentiable surrogate downstream classifier regularizer.
Key results
- On Adult non-private fair generation, CuTS reaches 82.1% XGB accuracy with 0.01 demographic parity distance, versus TabFairGAN 79.8% / 0.02 and DECAF 66.8% / 0.08.
- Under DP (epsilon=1) on Adult, CuTS achieves 80.9% accuracy and 0.01 demographic parity distance, versus Prefair Greedy 80.2% / 0.04.
- Downstream specification reduces balanced accuracy of predicting sex from 83.3% to 50.2% while retaining 84.4% task accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CuTS's fair-data specification instead of TabFairGAN | Adult synthetic data · Adult | XGB accuracy [%]: 79.8 ± 0.5 -> 82.1 ± 0.3 | Helped |
| Apply CuTS's fair-data specification instead of TabFairGAN | Adult synthetic data · Adult | Demographic parity distance (sex): 0.02 ± 0.01 -> 0.01 ± 0.01 | Helped |
| Apply CuTS fair-data specification with ε=1 instead of Prefair Greedy | Adult synthetic data · Adult (ε=1) | XGB accuracy [%]: 80.2 ± 0.4 -> 80.9 ± 0.3 | Helped |
| Apply CuTS fair-data specification with ε=1 instead of Prefair Greedy | Adult synthetic data · Adult (ε=1) | Demographic parity distance (sex): 0.04 ± 0.01 -> 0.01 ± 0.01 | Helped |
| Add fine-tuning on the logical constraint before rejection sampling (ablation) | CuTS · Adult logical constraints | XGB accuracy [%]: 82.9 ± 0.8 -> 84.7 ± 0.1; 84.5 ± 0.1 -> 84.8 ± 0.2; 85.1 ± 0.1 -> 85.0 ± 0.18 | Mixed |
| Set the average age to 30 using a statistical specification | CuTS on Adult · Adult | XGB accuracy [%]: 85.1 ± 0.16 -> 82.5 ± 0.76 | Hurt |
| Set male and female average ages equal using a statistical specification | CuTS on Adult · Adult | XGB accuracy [%]: 82.5 ± 0.76 -> 82.0 ± 0.50 | Hurt |
| Add a specification preventing prediction of sex | CuTS on Adult · Adult | Balanced accuracy for predicting sex: 83.3% -> 50.2% | Helped |
| Fine-tune on the constraint before rejection sampling (ablation) | CuTS · Health Heritage, Compas, and German logical constraints | XGB accuracy [%] | Mixed |
| Increase condition-feature entropy with a statistical specification | CuTS on Health Heritage · Health Heritage | Downstream accuracy: ≈ 1% | Hurt |
| Add fair-data, age-statistic, and implication specifications together | CuTS on Adult · Adult | XGB accuracy [%]: 85.1 ± 0.16 -> 81.6 ± 0.29 | Hurt |
Surprising findings
- Rejection sampling alone is insufficient for hard logical constraints; fine-tuning with relaxed constraints is needed for high performance.
- On Health Heritage, CuTS' fair generation drops accuracy considerably (70.9% vs 81.0% true data), indicating a notable accuracy-fairness trade-off.
Limitations
- Data must be discretized (uniform 32 buckets), which can hurt performance and is hard to avoid under DP.
- Uses simple three-way marginals for training without exploring better marginal selection.
- Supports only i.i.d. row-level constraints; constraints between pairs of data points are not covered.
Experimental setup
Datasets: Adult Health Heritage Compas German Credit
Benchmarks: Adult Health Heritage Compas German Credit
Metrics: XGBoost accuracy demographic parity distance constraint satisfaction rate balanced accuracy average age statistic
Baselines: TVAE CTGAN DECAF TabFairGAN Prefair AIM MST GReaT GEM
Builds on: AIM GEM Gumbel-softmax DL2
Closest papers by meaning
- Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning (NeurIPS 2026 · similarity 0.88)
- HARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy Protection (NeurIPS 2024 · similarity 0.86)
- Struct-Bench: A Benchmark for Differentially Private Structured Text Generation (NeurIPS 2025 · similarity 0.85)
- TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.85)
- TabSDS: a Lightweight, Fully Non-Parametric, and Model Free Approach for Generating Synthetic Tabular Data (ICML 2025 · similarity 0.85)
- Private Evolution Converges (NeurIPS 2025 · similarity 0.85)
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.