Research map / Graph learning / Graph self-supervised learning
CG-TGAN: Conditional Generative Adversarial Networks with Graph Neural Networks for Tabular Data Synthesizing
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 CG-TGAN, a conditional GAN whose generator and discriminator are graph convolutional networks over a fully connected column graph with a learnable adjacency matrix and a conditional node, trained with WGAN-GP plus node-level prediction, for tabular data synthesis.
Problem
Synthesizing realistic tabular data with mixed continuous, mixed-type and categorical columns, multimodal and imbalanced distributions, while preserving inter-column relationships and privacy; existing GAN models use fully connected or convolutional layers that are weakly inductive or order-dependen
Approach
Convert each tabular row into a fully connected graph where nodes are columns, use a learnable adjacency matrix initialized as J−I, add a conditional node connected to all columns in the discriminator, and train both networks with an additional node-level target-column prediction task alongside the graph-level WGAN-GP objective.
Key results
- On regression datasets, CG-TGAN achieved the smallest ML utility gap among GAN models on Abalone (MAE -0.1174 vs CTGAN -0.1765) and Insurance/King, per Table 1.
- Against diffusion models, CG-TGAN outperforms CoDi and is comparable to TabDDPM in ML utility, winning on 3 of 7 datasets (Insurance, King, Diabetes).
- CG-TGAN has higher DCR than TabDDPM on 6 of 7 datasets (e.g., Abalone 0.3418 vs 0.1569), suggesting better privacy protection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a conditional node (ablation) | CG-TGAN · tabular data synthesis; seven datasets | machine-learning utility | Mixed |
| Use residual connections (ablation) | CG-TGAN · tabular data synthesis; seven datasets | machine-learning utility | Helped |
| Initialize the adjacency matrix as JN − IN (ablation) | CG-TGAN · tabular data synthesis; seven datasets | machine-learning utility | Helped |
| Use three or four graph convolutional layers rather than one or two (ablation) | CG-TGAN · tabular data synthesis; seven datasets | machine-learning utility | Mixed |
| Use three graph convolutional layers instead of one or two (ablation) | CG-TGAN · Gesture | machine-learning utility | No clear effect |
| Replace the learned adjacency initialization with random initialization (ablation) | CG-TGAN · tabular data synthesis; seven datasets | machine-learning utility | Hurt |
| Use CG-TGAN instead of CoDi | tabular data synthesis · seven datasets | machine-learning utility | Helped |
| Use CG-TGAN instead of TabDDPM | tabular data synthesis · Insurance, King, and Diabetes | machine-learning utility | Mixed |
| Use CG-TGAN instead of TabDDPM | tabular data synthesis · Abalone, Adult, Diabetes, and Gesture | NNDR | Mixed |
| Use CG-TGAN instead of TabDDPM | tabular data synthesis · six of seven datasets | DCR | Helped |
Surprising findings
- Lower NNDR (closer to real records) is associated with higher synthetic data quality, so privacy and utility trade off across synthesizers.
- Ablations show a random-adjacency variant performs worse than the J−I initialization, but CG-TGAN's gains over GAN baselines are not uniform across all datasets (e.g., Adult accuracy gap is larger than CTAB-GAN).
Limitations
- Evaluation uses only seven datasets, several small (1000 or fewer rows), and single reported differences without variance.
- Privacy evaluated via NNDR and DCR heuristics rather than formal guarantees; NNDR results mixed versus GAN baselines.
- Comparison against diffusion models is limited to CoDi and TabDDPM, and the claim of outperforming GAN baselines is not uniform across all metrics.
Experimental setup
Datasets: Abalone King Insurance Adult Diabetes Wilt Gesture
Benchmarks: Abalone King Insurance Adult Diabetes Wilt Gesture
Metrics: MAPE MAE R2 accuracy AUC F1 score NNDR DCR
Baselines: DPGAN PATE-GAN CTGAN CTAB-GAN CTAB-GAN+ CoDi TabDDPM
Builds on: CTGAN CTAB-GAN Fi-GNN GCN WGAN-GP
Closest papers by meaning
- Modeling Tabular data using Conditional GAN (NeurIPS 2019 · similarity 0.92)
- GraphGAN: Graph Representation Learning With Generative Adversarial Nets (AAAI 2018 · similarity 0.88)
- TabGLM: Tabular Graph Language Model for Learning Transferable Representations Through Multi-Modal Consistency Minimization (AAAI 2025 · similarity 0.87)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.87)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.86)
- Geometric Generative Modeling with Noise-Conditioned Graph Networks (ICML 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.