[A-10] Mathematical Systems Science and its Applications

2:15 PM - 2:30 PM JST(5:15 AM - 5:30 AM UTC)

[A-10-05] Evaluation of Distribution Reproducibility for Unseen Data in Tabular Data Generation

〇Naoki Ikeda¹, Tomoki Oya¹, Mori Kurokawa¹ (1. KDDI Research, Inc.)

Keywords:

diffusion models,tabular data generation,conditional generation

Diffusion models have recently gained attention in tabular data generation for applications such as privacy preservation and data augmentation, due to their high capability in reproducing data distributions. Oya et al. [1] proposed a conditional generation framework that enables data generation conditioned on specific columns. While this approach allows the generation of unseen data through multiple condition settings, its performance has not yet been fully clarified. Therefore, in this study, we quantitatively evaluate the generation performance of synthetic data under unseen conditions using this conditional generation model.