IPL-UV / confound_it

Recovering latent confounders from high-dimensional proxy variables

This repository contains the code to replicate the experiments and examples from Recovering latent confounders from high-dimensional proxy variables

requirements

The scripts need R (for the simualtion experiments) and pyhton (for the real-world examples) installed.

The following R packages need to be installed:

• ggplot2
• reshape2
• pbapply
• optparse
• fastICA
• mdatools

The following python libraries need to be available:

• numpy
• pandas
• sklearn
• jax

reproduce the simulation experiments

To run the simulation experiments and produce the results we provide in the simulations/ folder:

• generate_data.R a script which generate the synthetic data (run Rscript generate_data.R --help to list all the options).
• run_methods.R a script which generate the synthetic data (run Rscript run_methods.R --help to list all the options).
• evaluate.R a script which evaluate the results (run Rscript evaluate.R --help to list all the options).
• get_plots.R a script to produce the final plots.

The exact simualtion and experiments can be reproduced with the commands in the experiments.sh file.

reproduce GD-PCF method and experiments

To apply the GD-PCF on the simulation experiments and produce the results we provide the gd-pcf/folder:

• funcs_LNC_lin.py functions that implement the GD-PCF method with linear assumptions
• funcs_LNC.py function that implement the GD-PCF method without assuming linearity.
• experiment.py script to apply GD-PCF on one dataset.
• slurm_script.py calls the experiment.py script for a dataset "job", parametrized by a slurm script that calls it
• processResults.py functions for gathering the results of GD-PCF for individual datasets which are stored as pickle files and obtaining performance measures.
• LNC_job.sh slurm script for running GD-PCF on different datasets in parallel