# TabDDPM: Modelling Tabular Data with Diffusion Models

**Authors:** Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, Artem Babenko  
**arXiv:** [2209.15421](https://arxiv.org/abs/2209.15421)  
**Date:** 2022-09-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.15421v1/

## Key points

- Denoising diffusion probabilistic models (DDPMs) are popular in generative modeling for various data modalities
- DDPMs have shown promise in computer vision, speech, NLP, and graph-like data
- Tabular data poses challenges due to its heterogeneity
- TabDDPM is a diffusion model specifically designed for tabular data
- TabDDPM can handle any feature type present in the tabular dataset
- TabDDPM outperforms GANs and VAEs on benchmark datasets
- TabDDPM is suitable for privacy-oriented setups where original datapoints cannot be publicly shared

## Summary

In the field of generative modeling, denoising diffusion probabilistic models (DDPMs) have emerged as a leading paradigm for various data modalities. While these models have gained significant popularity in computer vision, they have also shown promise in other domains such as speech, natural language processing (NLP), and graph-like data. This study aims to explore the potential advantages of using diffusion models for general tabular problems. Tabular data poses unique challenges for accurate modeling due to its inherent heterogeneity. Each datapoint in a tabular dataset is typically represented by a vector of features that can vary widely in nature. Some features may be continuous, while others may be discrete. This diversity makes it difficult to develop effective models that can capture the underlying patterns and generate realistic samples. To address this issue, the researchers propose TabDDPM, a diffusion model specifically designed for tabular data. TabDDPM is a universal model that can be applied to any tabular dataset regardless of the feature types present. It leverages the framework of diffusion models to effectively model and generate samples from heterogeneous tabular data. The performance of TabDDPM is extensively evaluated on a wide range of benchmark datasets. The results demonstrate its superiority over existing alternatives such as generative adversarial networks (GANs) and variational autoencoders (VAEs). This finding aligns with the advantage observed in diffusion models across different fields. Furthermore, the study highlights an additional benefit of TabDDPM: its eligibility for privacy-oriented setups where original datapoints cannot be publicly shared. This feature makes TabDDPM suitable for scenarios where preserving data privacy is crucial. Overall, this research contributes to advancing generative modeling techniques by introducing TabDDPM as an effective solution for modeling tabular data with heterogeneous features.