NeurIPS 2024

Official implementation of "ClavaDDPM:Multi-relational Data Synthesis with Cluster-guided Diffusion Models" https://arxiv.org/pdf/2405.17724

To run a train-sampling pipeline:

python complex_pipeline --config_path configs/movie_lens.json

To use customized datasets: See complex_data/california as an example.

1. Save all tables as .csv files. All id columns should be named <column_name>_id.
2. Create a dataset_meta.json, in which tables should be manually created to specify all foreign key relationships in a multi-table dataset.
3. Create a relation_order in dataset_meta.json, which specifies the topological order of the multi-table dataset. The function topological_sort in preprocess_utils.py helps create it.
4. Create a domain file for each table, id columns excluded.

Other datasets

(CCS, Instacart 05, Berka): https://drive.google.com/file/d/1H_IGsL7JuCJYlq_6TOmdIZ4-q-wFafCL/view?usp=sharing

Model checkpoints

Some trained model checkpoints are provided for the convenience of development: https://drive.google.com/file/d/1dmSdi-5Wc3uad5QPrIu2ihysAIOIUUS0/view?usp=sharing, this includes model checkpoints, cluster checkpoints, synthetic data, and training log.