disjoint-generation 1.1

Exploring the Effects of Disjoint Generation of Synthetic Tabular Data.

Disjoint Generative Models

Disjoint Generative Models (DGMs) is a framework for generating synthetic data by distributing the generation of different attributes to different generative models. DGMs unlock mixed model generation, allowing the user to choose ``correct tool for the correct job'' and infers increased privacy by not having a single model that has access to all the data.

The library provides a simple API for generating synthetic data using a variety of generative models and joining strategies. The library has access to a variety of generative model backends namely SynthCity, DataSynthesizer and Synthpop, but additional backends can be added in the adapters module. Similarly several methods for joining are available for combining the generated data, and more can be added in the joining strategies module.

Installation

We recommend cloning and working with the repository directly due to challenging dependencies, but the library is also available on PyPI. To install the library, run the following command:

pip install disjoint-generation

One of the generative model backends "synthpop" requires a working R installation on the system. Access is handled through subprocess to run an Rscript command, so make sure that the Rscript command works in the terminal.

Tutorial and Codebooks

Below is codebooks that can be used to replicate the results shown in the paper.

Link	Description	Fig. refs.
Tutorial	A simple tutorial on how to use the library	NA
Codebook 1	Introductionary experiments, random joining, incresing number of partitions	Fig. 3, 10
Codebook 2	High-dimensional dataset example vith validation, correlated partitions study	Fig. 4, 5, 6, 13
Codebook 3	Mixed-model generation and combinatorics	Fig. 7, 8, Tab. 2, 3
Codebook 4	Study of the joining validator model, optimisation and calibration	Fig. 11, 12, 14
---	---	---
TabDiff Baseline	We compare to TabDiff, a state-of-the-art method for synthetic data generation. Our code and adapters for making this baseline work is provided here.	Tab. 3, Fig. 8

Additional examples for how to use the library can be seen in the documentation in the source code folder.

Requirements

The codebase requires Python ~3.10 (we use version 3.10.11) and the following packages:

• numpy ~= 1.26
• pandas ~= 2.2.3
• scipy ~= 1.12
• scikit-learn ~= 1.5
• synthcity >= 0.2.11
• DataSynthesizer ~= 0.1.13
• pyod >= 2.0

Because Synthcity sometimes causes compatibility issues we also provide an environment.txt file with a pip freeze of a working installation, which can be downloaded and installed using pip install -r environment.txt and should work on Python ~3.10.11.

Additonally, the synthpop generative model is accessed through R (we used version 4.1.2), and requires the following R packages:

• synthpop ~= 1.8.0

Citing

If you use our library in your work, you can reference us by citing our paper:

@misc{Lautrup2025,
      title={Disjoint Generative Models}, 
      author={Anton Danholt Lautrup and Muhammad Rajabinasab and Tobias Hyrup and Arthur Zimek and Peter Schneider-Kamp},
      year={2025},
      eprint={2507.19700},
      archivePrefix={arXiv},
}