Latent Diffusion Models with Correlation Loss for Tabular Data Generation
Qingyu Zhang · 2025
The generation of high-quality synthetic tabular data has emerged as a crucial area of research, driven by its wide-ranging applications in data augmentation and privacy-preserving data sharing. With the rapid development of deep generative models, particularly diffusion models, the generation of tabular data via such models has garnered significant attention. In this paper, we propose a novel diffusion-based tabular data generation model, which utilizes autoencoders to map tabular data into a latent space without making any assumptions about the latent space distribution. The latent space distribution is subsequently learned through a diffusion process. To enhance the model's ability to capture attribute correlations and generate higher-quality tabular data, we explicitly model these correlations and introduce a correlation loss. Additionally, the reconstruction output of the autoencoder is incorporated as a condition for the latent diffusion, stabilizing the training and improving overall performance. Extensive experiments on real datasets demonstrate that our proposed method outperforms existing methods in terms of both statistical metrics and machine learning efficacy.