limorigu / ComplexCauses

Code for the ICML 2021 paper "Operationalizing Complex Causes: A Pragmatic View of Mediation"

Operationalizing Complex Causes: A Pragmatic View of Mediation

This folder contains code to reproduce results for the ICML 2021 paper. Please note that certain files referenced by some of the scripts included.

Prereqs

The following will need to be downloaded and generated for the code to run:

Python_Img_Humor/data/

1. you will need glove.6B, downloaded from https://nlp.stanford.edu/projects/glove/.
2. (optional) one could also use GoogleNews-vectors-negative300.bin.gz, download from https://code.google.com/archive/p/word2vec/, under 'Pre-trained word and phrase vectors'.
3. you will need semeval-2020-task-7-dataset/*, downloaded from https://www.cs.rochester.edu/u/nhossain/humicroedit.html, 'Full dataset release'.

Python_Img_Humor/data/ImgPretSim/diffY and Python_Img_Humor/data/Humicroedit/diffY

To reproduce Figure 8 you will need to generate the datasets with different Y configs, by running the last cell in the notebooks Python_Img_Humor/data_generation_notebooks/ImgPretSim/ImgPretSim.ipynb and Python_Img_Humor/data_generation_notebooks/Humicroedit/Humicroedit.ipynb.

Image Perturbation

Code can be found in Python_Img_Humor/src and data can be found in Python_Img_Humor/data/ImgPertSim. Description of dataset construction is in the Jupyter notebook Python_Img_Humor/data_generation_notebooks/ImgPertSim/ImgPertSim.ipynb. The code is written in Pytorch. To reproduce results follow:

1. Make sure requirements are installed, i.e. pip install requirements.txt (suggested: use GPU if available). Alternatively, use the environment.yml file, and create a virtual environment with

conda env create -f environment.yml
conda activate ComplexCauses

2. Configurations for the Image Preturbation setup are indicated in Python_Img_Humor/src/configs/causal_effect_config.yml. Make sure the ones under "ImgPertSim" headlines are uncommented (and make sure to comment out Humicroedit configs).
3. Call the command "python main.py"
4. Results will appear in Python_Img_Humor/out/viz/ImgPertSim. MSE_vs_Y_labels_perc.png will correspond to Figure 5.
5. To reproduce Figure 6, first consider the parameters of fitted model, saved in df_params_PertImgSim.csv. Next, call the conditional independence test in R_cond_ind_test/ci_test_img.R on the files Python_Img_Humor/out/viz/ImgPertSim/residuals_nested=True_PertImgSim.csv and Python_Img_Humor/out/viz/ImgPertSim/residuals_nested=False_PertImgSim.csv. nested=True matches the model \Phi ~ Z, and nested=False matches the model \Phi ~ W,Z. Results will be saved in R_cond_ind_test/ci_test_results_img.csv

Humor edits

Code can be found in Python_Img_Humor/src and data can be found in Python_Img_Humor/data/Humicroedit. Description of dataset construction is in the Jupyter notebook Python_Img_Humor/data_generation_notebooks/Humicroedit/Humicroedit.ipynb. The code is written in Pytorch. To reproduce results follow:

1. Make sure requirements are installed, i.e. pip install requirements.txt (suggested: use GPU if available)
2. Configurations for the Image Preturbation setup are indicated in Python_Img_Humor/src/configs/causal_effect_config.yml. Make sure the ones under "Humicroedit" headlines are uncommented (and make sure to comment out ImgPert configs).
3. Call the command "python main.py"
4. Results will appear in Python_Img_Humor/out/viz/Humicroedit. MSE_vs_Y_labels_perc.png will correspond to Figure 5.
5. To reproduce Figure 6, first consider the parameters of fitted model, saved in df_params_Humicroedit.csv. Next, call the conditional independence test in R_cond_ind_test/ci_test_humor.R . The paths to the correct files is already prespecified in the file, but for completeness, it will use the files Python_Img_Humor/out/viz/Humicroedit/residuals_nested=True_PertImgSim.csv and Python_Img_Humor/out/viz/Humicroedit/residuals_nested=False_Humicroedit.csv. nested=True matches the model \Phi ~ Z, and nested=False matches the model \Phi ~ W,Z. Results will be saved in R_cond_ind_test/ci_test_results_humor.csv

Gene Knockouts

Code can be found in R_genomics/. The data was too big to include herein, but instructions for download are included. To reproduce results follow:

1. Your working directory will need the following (already created in this submission) directories ready, where model objects and simulated data will be stored: "genie3_models" (this will contain model objects corresponding to the 4177 structural equations of the GENIE3 model); "simulations" (this will contain data for Z, W, X, Phi, Phi_hat, and Y); and "phi_models" (this will contain model objects for the 168 E[\Phi | Z, W] regressions).
2. Download net3_expression_data.tsv from http://dreamchallenges.org/project/dream-5-network-inference-challenge/ and place it in your working directory.
3. Call "Rscript pkgs.R" to install all requisite libraries.
4. Call "Rscript fit_genie3.R" to train GENIE3 on the E. Coli dataset and export model objects (this takes ~8 hrs).
5. Call "Rscript e_coli_sim.R" to simulate Z, X, and Phi (this takes ~1 hr).
6. Call "Rscript phi_models.R" to simulate Phi_hat (this takes ~4 hrs).
7. Call "Rscript y_model.R" to simulate Y and fit a lasso model for E[Y | Phi_hat]
8. Call "Rscript baselines.R" to fit a series of baseline models for E[Y | Z, W] with increasing amounts of data.
9. Call "Rscript discover_m.R" to perform association tests for Phi and W, thereby completing the mediator discovery pipeline.