syntherela 0.0.1

SyntheRela - Synthetic Relational Data Generation Benchmark

SyntheRela - Synthetic Relational Data Generation Benchmark

Our paper Benchmarking the Fidelity and Utility of Synthetic Relational Data is available on arxiv.

Installation

To install only the benchmark package, run the following command:

pip install .

Replicating the paper's results

We divide the reproducibility of the experiments into two parts: the generation of synthetic data and the evaluation of the generated data. The following sections describe how to reproduce the experiments for each part.

To reproduce some of the figures the synthetic data needs to be downloaded first. The tables can be reproduced with the results provided in the repository or by re-running the benchmark.

First, create a .env file in the root of the project with the path to the root of the project. Copy .env.example, rename it to .env and update the path.

Download synthetic data and results

The data and results can be downloaded and extracted with the below script, or are available on google drive here.

conda activate reproduce_benchmark
./experiments/reproducibility/download_data_and_results.sh

Evaluation of synthetic data

To run the benchmark and get the results of the metrics, run:

conda activate reproduce_benchmark
./experiments/reproducibility/evaluate_relational.sh

./experiments/reproducibility/evaluate_tabular.sh

./experiments/reproducibility/evaluate_utility.sh

Generation of synthetic data

Depending on the synthetic data generation method a separate pythone environment is needed. The instruction for installing the required environment for each method is provided in docs/INSTALLATION.md.

After installing the required environment, the synthetic data can be generated by running the following commands:

conda activate reproduce_benchmark
./experiments/reproducibility/generation/generate_sdv.sh

conda activate rctgan
./experiments/reproducibility/generation/generate_rctgan.sh

conda activate realtabformer
./experiments/reproducibility/generation/generate_realtabformer.sh

conda activate tabular
./experiments/reproducibility/generation/generate_tabular.sh

conda activate gretel
python experiments/generation/gretel/generate_gretel.py --connection-uid  <connection-uid> --model lstm
python experiments/generation/gretel/generate_gretel.py --connection-uid  <connection-uid> --model actgan

cd experiments/generation/clavaddpm
./generate_clavaddpm.sh <dataset-name> <real-data-path> <synthetic-data-path>  

To generate data with MOSTLYAI, insructions are provided in experiments/generation/mostlyai/README.md. 
Further instructions for GRETELAI are provided in experiments/generation/gretel/README.md.

Visualising Results

To visualize results, after running the benchmark you can run the below script. The figures will be saved to results/figures/:

conda activate reproduce_benchmark
./experiments/reproducibility/generate_figures.sh

Reproducing Tables

To reproduce the tables you can run the below script. The tables will be saved as .tex files in results/tables/:

conda activate reproduce_benchmark
./experiments/reproducibility/generate_tables.sh

Adding a new metric

The documentation for adding a new metric can be found in docs/ADDING_A_METRIC.md.

Synthetic Data Methods

Open Source Methods

• SDV: The Synthetic Data Vault
• RCTGAN: Row Conditional-TGAN for Generating Synthetic Relational Databases
• REaLTabFormer: Generating Realistic Relational and Tabular Data using Transformers
• ClavaDDPM: Multi-relational Data Synthesis with Cluster-guided Diffusion Models
• IRG: Generating Synthetic Relational Databases using GANs
• Generating Realistic Synthetic Relational Data through Graph Variational Autoencoders*
• Generative Modeling of Complex Data*
• BayesM2M & NeuralM2M: Synthetic Data Generation of Many-to-Many Datasets via Random Graph Generation*

* Denotes the method does not have a public implementation available.

Commercial Providers

A list of commercial synthetic relational data providers is available in docs/SYNTHETIC_DATA_TOOLS.md.

Conflicts of Interest

The authors declare no conflict of interest and are not associated with any of the evaluated commercial synthetic data providers.

Citation

If you use SyntheRela in your work, please cite our paper:

@misc{hudovernik2024benchmarkingsyntheticrelationaldata,
      title={Benchmarking the Fidelity and Utility of Synthetic Relational Data}, 
      author={Valter Hudovernik and Martin Jurkovič and Erik Štrumbelj},
      year={2024},
      eprint={2410.03411},
      archivePrefix={arXiv},
      primaryClass={cs.DB},
      url={https://arxiv.org/abs/2410.03411}, 
}