This repository contains code for our paper accepted in ICML Workshop 2022 on Spurious Correlations, Invariance and Stability (SCIS): Causal Discovery using Model Invariance via Knockoffs by Wasim Ahmad, Maha Shadaydeh and Joachim Denzler.

• The work can be cited using below citation information.

@inproceedings{
ahmad2022causal,
title={Causal Discovery using Model Invariance through Knockoff Interventions},
author={Wasim Ahmad and Maha Shadaydeh and Joachim Denzler},
booktitle={ICML 2022: Workshop on Spurious Correlations, Invariance and Stability},
year={2022},
url={https://openreview.net/forum?id=OcNeMVbIdCF}
}

We discover full causal graph in multivariate nonlinear systems by testing model invariance against Knockoffs-based interventional environments:

1. First we train deep network $f_i$ using data from observational environment $E_i$.
2. Then we expose the model to Knockoffs-based interventional environments $E_k$.
3. For each pair variables ${z_i, z_j}$ in nonlinear system, we test model invariance across environments.
4. We perform KS test over distribution $R_i$, $R_k$ of model residuals in various environments. Our NULL hypothesis is that variable $z_i$ does not cause $z_j$, $H_0$: $R_i$ = $R_k$, else the alternate hypothesis $H_1$: $R_i$ != $R_k$ is accepted.

We test our method on synthetic as well as real data which can be found under datasets/ directory. The synthetic data is generated using file src/synthetic_dataset.py. The real data we used is average daily discharges of rivers in the upper Danube basin, measurements of which are made available by the Bavarian Environmental Agency at https://www.gkd.bayern.de.

src/main.py is our main file, where we model multivariate non-linear data using deep networks.

• src/deepcause.py for actual and counterfactual outcome generation using interventions.
• src/functions.py for preprocessing and methods.
• src/knockoffs.py generate knockoffs of the original variables.
• src/daignostics.py to determine the goodness of the generated knockoff copies.
• src/cdmi_examples.ipynb for testing various examples for causal discovery.
• DeepKnockoffs/ contains the knockoffs generation methods.
• datasets/ contains the generated synthetic data and real dataset.
• model/ contains trained models that we used for different datasets.

We also extend CDMI to discover causal graph in non-stationary time series. To do that, first we segment multivariate non-stationary time series based on their dynamics using Regime Identification method and perform regime-oriented causal discovery.

requirements.txt contains all the packages that are related to the project. To install them, simply create a new conda environment (use python=3.8.15) and type

pip install -r requirements.txt

This work is funded by the Carl Zeiss Foundation within the scope of the program line "Breakthroughs: Exploring Intelligent Systems" for "Digitization — explore the basics, use applications" and the DFG grant SH 1682/1-1.