Research map / Image generation / Autoregressive and latent generation
Beyond the convexity assumption: Realistic tabular data generation under quantifier-free real linear constraints
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 the Disjunctive Refinement Layer (DRL), a differentiable output layer that compiles QFLRA constraints via a generalised Fourier-Motzkin/cutting-planes variable elimination so that tabular DGM samples always satisfy the constraints while improving ML efficacy.
Problem
Deep generative models for tabular data frequently produce samples that violate background knowledge expressed as constraints, and prior constraint layers only handle convex linear inequalities.
Approach
A layer compiled from quantifier-free linear real arithmetic formulas (disjunctions of linear inequalities) via a CP resolution-based variable elimination; at inference it sequentially projects each feature to the nearest feasible boundary, guaranteeing satisfaction in one forward pass and optimality w.r.t. a variable ordering.
Key results
- DRL reduces constraint violation rate to 0% for all five DGMs across all five datasets, versus up to 100% for unconstrained DGMs.
- DGM+DRL improves F1 in 19 of 20 classification cases, with the largest gain of 21.4% on GOGGLE for CCS.
- DGM+DRL outperforms DGM+LL in 13 of 20 F1 comparisons, and sampling is faster than rejection sampling (e.g., 0.22s vs 0.37s on URL).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DRL to DGMs | Five tabular DGMs · URL, CCS, LCLD, Heloc, and House | constraint violation rate: 0.0\msmall{\pm0.0} | Helped |
| Add DRL to GOGGLE | GOGGLE · CCS | F1: 0.039 -> 0.253 | Helped |
| Add DRL to GOGGLE | GOGGLE · CCS | weighted F1: 0.076 -> 0.281 | Helped |
| Add DRL to CTGAN | CTGAN · CCS | AUC: 0.914 -> 0.955 | Helped |
| Use rejection sampling instead of DRL | Tabular DGMs · Classification datasets | machine learning efficacy | Hurt |
| Use DRL instead of a linear layer | DGM+LL · Classification datasets | F1 | Mixed |
| Use DRL instead of a linear layer | DGM+LL · Classification datasets | weighted F1 | Mixed |
| Use DRL instead of a linear layer | DGM+LL · Classification datasets | AUC | Mixed |
| Add DRL to DGMs | Five tabular DGMs · Classification datasets | F1: 21.4% | Helped |
| Add DRL to DGMs | Five tabular DGMs · Classification datasets | weighted F1: 20.5% | Helped |
| Add DRL to DGMs | Five tabular DGMs · Classification datasets | AUC: 20.9% | Helped |
Surprising findings
- Rejection sampling to enforce constraints often reduces ML efficacy compared with unconstrained DGMs.
- Unconstrained DGMs violate constraints in 13 of 25 DGM-dataset scenarios by over 50%, with five at 100%.
Limitations
- Compilation via variable elimination can produce a non-polynomial number of constraints.
- Improvements over the linear layer are more modest, and DGM+DRL beats DGM+LL in only 13/20 F1 cases.
- Sampling is slower than unconstrained DGMs and the variable ordering affects results.
Experimental setup
Datasets: URL CCS LCLD Heloc House
Benchmarks: URL CCS LCLD Heloc House
Metrics: constraint violation rate samplewise constraints violation coverage constraint violation coverage F1 weighted F1 AUC MAE RMSE sample generation time
Baselines: WGAN TableGAN CTGAN TVAE GOGGLE rejection sampling linear layer (LL)
Builds on: Linear Layer for tabular constraints Fourier-Motzkin elimination cutting planes
Closest papers by meaning
- How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.89)
- TabSDS: a Lightweight, Fully Non-Parametric, and Model Free Approach for Generating Synthetic Tabular Data (ICML 2025 · similarity 0.85)
- Tabular Foundation Model for Generative Modelling (NeurIPS 2026 · similarity 0.84)
- Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning (NeurIPS 2026 · similarity 0.84)
- TabM: Advancing tabular deep learning with parameter-efficient ensembling (ICLR 2025 · similarity 0.84)
- PTaRL: Prototype-based Tabular Representation Learning via Space Calibration (ICLR 2024 · similarity 0.83)
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.