Research map / Label-efficient and robust learning / Recognition and long-tail
A Learnable Discrete-Prior Fusion Autoencoder with Contrastive Learning for Tabular Data Synthesis
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 GTCoder, a transformer-fusion autoencoder with a learnable Gaussian-mixture discrete prior and contrastive loss for synthesizing tabular data with mixed discrete and continuous features.
Problem
Synthesizing privacy-preserving tabular data that contains both discrete and continuous features, where existing generators treat the two modalities separately and fail to capture their correlations.
Approach
Fuses continuous and discrete tokens with a transformer [CLS]-based encoder, models discrete features with a Gaussian mixture (GMVAE-style) learnable prior whose embeddings are updated during training, and adds a contrastive loss that pulls together discrete-feature embeddings with similar distributions and pushes dissimilar ones apart.
Key results
- GTCoder achieves the best ma-F1 and we-F1 on most real datasets, e.g. Agaricus-lepiota 0.953 ma-F1 vs. 0.92 for the best baseline (Table 1 / Table 3).
- Ablation on Pregnancy shows the full model (CL + learnable prior + transformer) reaches 0.657 ma-F1 and 0.830 we-F1 versus 0.573/0.710 without components (Table 3).
- On simulated datasets GTCoder obtains the best Lsyn on Grid, Gridr and Ring relative to baselines in most cases, though CTGAN is slightly better on some minority metrics (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GTCoder instead of the strongest listed baseline | Tabular data synthesis · Adult | ma-F1: 0.69 -> 0.70 | Helped |
| Use GTCoder instead of the strongest listed baseline | Tabular data synthesis · Ionosphere | ma-F1: 0.51 -> 0.60 | Helped |
| Use GTCoder instead of the strongest listed baseline | Tabular data synthesis · Agaricus-lepiota | ma-F1: 0.92 -> 0.95 | Helped |
| Add contrastive learning to the model with learnable prior and transformer fusion (ablation) | GTCoder · Ionosphere | ma-F1: 0.562 -> 0.603 | Helped |
| Add contrastive learning to the model with learnable prior and transformer fusion (ablation) | GTCoder · Agaricus-lepiota | ma-F1: 0.936 -> 0.953 | Helped |
| Add contrastive learning to the model with learnable prior and transformer fusion (ablation) | GTCoder · Pregnancy | ma-F1: 0.643 -> 0.657 | Helped |
| Add the learnable prior to the model with transformer fusion (ablation) | GTCoder · Ionosphere | ma-F1: 0.483 -> 0.562 | Helped |
| Add the learnable prior to the model with transformer fusion (ablation) | GTCoder · Agaricus-lepiota | ma-F1: 0.896 -> 0.936 | Helped |
| Add the learnable prior to the model with transformer fusion (ablation) | GTCoder · Pregnancy | ma-F1: 0.596 -> 0.643 | Helped |
| Add transformer fusion to the model without contrastive learning or learnable prior (ablation) | GTCoder · Ionosphere | ma-F1: 0.479 -> 0.483 | Helped |
| Add transformer fusion to the model without contrastive learning or learnable prior (ablation) | GTCoder · Agaricus-lepiota | ma-F1: 0.882 -> 0.896 | Helped |
| Add transformer fusion to the model without contrastive learning or learnable prior (ablation) | GTCoder · Pregnancy | ma-F1: 0.573 -> 0.596 | Helped |
Surprising findings
- GTCoder slightly underperforms CTGAN on Abalone and Sonar, and the authors note it lacks strength on simulated datasets with few discrete features.
Limitations
- Weaker on datasets with few discrete features (Abalone, Sonar, simulated sets) where CTGAN is competitive.
- Ablation indicates the contrastive loss alone gives limited gains without the learnable prior.
- Evaluation relies on downstream classifier/regressor utility rather than privacy metrics.
Experimental setup
Datasets: Adult News Abalone Ionosphere Agaricus-lepiota Sonar Pregnancy grid gridr ring
Benchmarks: Adult News Abalone Ionosphere Agaricus-lepiota Sonar Pregnancy grid gridr ring
Metrics: macro-F1 weighted-F1 r2 likelihood fitness (Lsyn, Ltest)
Baselines: CTGAN Bi-discriminator GAN WGAN TVAE LVAE GMVAE
Builds on: GMVAE tabular transformer (FT-Transformer style) CTGAN contrastive learning for tabular data (Bai et al. 2022)
Closest papers by meaning
- SwitchTab: Switched Autoencoders Are Effective Tabular Learners (AAAI 2024 · similarity 0.86)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.85)
- ICE-T: Interactions-aware Cross-column Contrastive Embedding for Heterogeneous Tabular Datasets (AAAI 2025 · similarity 0.85)
- Unity by Diversity: Improved Representation Learning for Multimodal VAEs (NeurIPS 2024 · similarity 0.84)
- Tabular Foundation Model for Generative Modelling (NeurIPS 2026 · similarity 0.84)
- Multi-modal contrastive learning adapts to intrinsic dimensions of shared latent variables (NeurIPS 2025 · similarity 0.84)
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.