Continuous Diffusion for Mixed-Type Tabular Data

---

Changelog

• improved sampling efficiency by caching all normalized embeddings once before generation
• switched to uv for easier replication of environment (use uv sync to install required packages)

---

This is an easy-to-use wrapper around the training and sampling procedures of the CDTD model.

Please see main.ipynb for an illustration of the CDTD on the adult dataset.

Paper: https://arxiv.org/abs/2312.10431 (published in ICLR 2025)

Full Replication Code: https://github.com/muellermarkus/cdtd

Install Instructions

Run uv sync.

Citation

@inproceedings{mueller2025cdtd,
  title = {Continuous {{Diffusion}} for {{Mixed-Type Tabular Data}}},
  author = {Mueller, Markus and Gruber, Kathrin and Fok, Dennis},
  booktitle = {The Thirteenth International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/pdf?id=QPtoBPn4lZ}
}