[2K5-GS-2] Machine learning

[2K5-GS-2-03] Generating a Wide Variety of Categorical Data Using Diffusion Models

〇Masane Fuchi¹, Amar Zanashir², Hiroto Minami², Tomohiro Takagi¹ (1. Meiji University, 2. LAC Co., Ltd.)

Keywords:Tabular Data Generation, Diffusion Models, Categorical Data, Deep Learning

Diffusion Models, which have been frequently researched in the field of Computer Vision as a method that outperforms GANs, are not confined to that field, but are spreading to other fields as well. TabDDPM using Diffusion Models have also been proposed for tabular data generation, and its authors claimed it can generate highly accurate data. However, TabDDPM tends to generate similar data as the number of such categories increases, resulting in learning filures because it handles categorical data as one-hot vectors. In order to overcome that problem, in this paper, we propose Table BD (Table Bit Diffusion) by incorporating Bit Diffusion preprocessing method. In our experiments, Table BD can generate data with a larger number of categories than TabDDPM.