Trans-CTGAN: A Transformer-Enhanced GAN Model for Correlation-Preserving Tabular Data Generation

High-fidelity synthetic tabular data generation is crucial, given that downstream tasks and data analysis typically require large amounts of tabular data. However, existing methods struggle to model complex inter-feature correlations in tabular datasets (e.g., customer behavior logs or healthcare records). To address this challenge, we propose Trans-CTGAN, an improved GAN model integrating Transformer encoder. This model introduces multi-head attention mechanisms in both generator and discriminator to optimize global dependency modeling, enabling comprehensive capture of feature interactions. Extensive experiments on public datasets demonstrate superior performance: a 26.25 % average improvement in the $\mathbf{L 2}$ distance of correlation matrices compared to baselines. This work provides a robust solution for generating high-fidelity synthetic tabular data with preserved feature correlations, offering both theoretical and practical value for data-sensitive applications.