RelationalCausalModels

Code for the experiments in the paper Relational Structural Causal Models by Adiba Ejaz and Elias Bareinboim, presented at ICML 2026.

Setup

python -m pip install -r requirements.txt

Training scripts default to GPU execution.

Repository layout

• experiments/: paper experiment entrypoints and baselines.
• plotting/: plotting and graph-visualization scripts.
• tools/: utility scripts for graph generation and support-trial search.
• relational/: relational modeling components contributed in this repository.
• NCMCounterfactuals/: vendored base NCM/SCM implementation from CausalAILab/NCMCounterfactuals, included so the experiment code is self-contained. The vendored code remains under its original MIT License; see NCMCounterfactuals/LICENSE and NCMCounterfactuals/NOTICE.md.
• out/: generated results, logs, locks, trial data artifacts, and plots. This directory is ignored by git.

Reproducing paper experiments

The paper experiments write generated data, logs, results, and plots under out/. Relational graph files are included under NCMCounterfactuals/dat/cg/relational/.

For exact data-generation, training, baseline, and plotting commands, see docs/reproduction.md.

Implementation notes:

• docs/relational-components.md: maps the relational CTM data-generation code and the RNCM training modules.

Citation

@inproceedings{
ejaz2026relational,
title={Relational Structural Causal Models},
author={Adiba Ejaz and Elias Bareinboim},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=WTBaZHtIra}
}