Bridging stability and utility in synthetic tabular data: The sTableGAN model

BOZKURT M. H.

Knowledge-Based Systems, cilt.336, 2026 (SCI-Expanded, Scopus) identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 336
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.knosys.2026.115303
  • Dergi Adı: Knowledge-Based Systems
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Library, Information Science & Technology Abstracts (LISTA)
  • Anahtar Kelimeler: Data augmentation, Generative adversarial networks, Machine learning utility, Medical data privacy, Mode collapse, Statistical similarity, Synthetic data generation, Tabular data
  • Karadeniz Teknik Üniversitesi Adresli: Evet

Özet

Synthetic tabular data is increasingly used to address privacy, scarcity, and regulatory constraints in healthcare. Several GAN-based approaches experience training instability, mode collapse, and inadequate preservation of minority classes. We propose sTableGAN, a practical variant of TableGAN that improves stability without explicit gradient penalties or feature matching. The model uses binary cross‑entropy loss, asymmetric discriminator updates, and label smoothing, while retaining the tabular grid pipeline and employing Swish activations in the discriminator. We evaluate sTableGAN against CTGAN, T‑VAE‑GAN, CTAB‑GAN, CGAN, and classic TableGAN on three medical datasets (Mental Health, Gallstone, DIA). Evaluation covers statistical similarity (Jensen–Shannon Divergence and Kolmogorov–Smirnov tests), machine‑learning utility (Train on Synthetic, Test on Real / Train on Real Test on Synthetic), privacy (discriminative and membership‑inference AUC), and diversity (class coverage and class‑balance KL). sTableGAN delivers consistently lower distributional divergence than classic TableGAN and strong KS outcomes. It maintains full class coverage and near‑zero class‑balance error, indicating the absence of mode collapse. Utility remains competitive across datasets under standard protocols. Privacy metrics are near random, and nearest‑neighbor analyses show no evidence of near‑copy generation. These results indicate that sTableGAN offers a balanced trade‑off between fidelity, utility, and privacy for sensitive tabular data. The approach provides a reliable and scalable alternative to existing GAN baselines for downstream medical applications.