Identifying Linearly-Mixed Causal Representations from Multi-Node Interventions

This is the official accompanying code repository for the paper Identifying Linearly-Mixed Causal Representations from Multi-Node Interventions by Simon Bing, Urmi Ninad, Jonas Wahl and Jakob Runge.

Requirements

This code was developed with Python 3.10 and PyTorch 2.0.1. Install the required dependencies by running

conda env create -f environment.yml

and activate the environment by running

conda activate multi_node_crl

Experiments

Our experiments can all be reproduced from a single script. To do so, run:

python experiment_main.py

The most important flags are:

• --models to select which models to compare (choose from ours,icrl,ica).
• --scm_id to select which underlying SCM model is used.
• --d to pass the number of nodes in the underlying model to generate the data.