CSI-VAE

Official implementation of the paper "Structured Learning of Compositional Sequential Interventions" (NeurIPS 2024).

This repository contains the code for training and evaluating the CSI-VAE model, which allows us to build predictive models that can identify and learn the effects of intervention combinations in sequential settings, particularly in sparse data regimes where only limited combinations are jointly observed.

• Anaconda or Miniconda
• Python 3.8+
• Git

1. Clone the repository:

git clone https://github.com/jialin-yu/CSI-VAE.git
cd CSI-VAE

2. Create and activate the conda environment:

conda env create -f environment.yml -n csi-vae
conda activate csi-vae

1. Generate synthetic data:

   • Run synthetic/simulator-clean.ipynb
   • The simulator creates the necessary datasets for all synthetic experiments

2. Train and evaluate models:

   • Execute notebooks in the synthetic/ directory for different model variants
   • For conformal prediction experiments, use synthetic/conformal-prediction.ipynb

Note: Ensure consistent random seeds between the simulator and training notebooks for reproducibility.

1. Data Preparation:

   • Run preprocess.ipynb to process the raw data (download link provided in notebook)
   • Alternatively, use load_data.ipynb to load pre-processed datasets
   • Generate experimental data using simulator.ipynb

2. Training and Evaluation:

   • Navigate to spotify/ directory
   • Follow individual notebook instructions for specific experiments

Use visualisation.ipynb to reproduce figures and visualize experimental results.

If you find this code useful for your research, please cite our paper:

@article{yu2024structured,
  title={Structured Learning of Compositional Sequential Interventions},
  author={Yu, Jialin and Koukorinis, Andreas and Colombo, Nicol{\`o} and Zhu, Yuchen and Silva, Ricardo},
  journal={arXiv preprint arXiv:2406.05745},
  year={2024}
}

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

For questions or issues, please open a GitHub issue or contact one of the authors (jialin.yu@ucl.ac.uk).