Active Learning for Optimal Interventions

Code for paper: Active learning for optimal intervention design in causal models (Nature Machine Intelligence, 2023)

arXiv link: https://arxiv.org/abs/2209.04744

Installation

Follow the two steps illustrated below

1. create a conda environment using environment.yaml (all dependencies are included; whole process takes about 5 min):

conda env create -f environment.yml

2. install the current package in editable mode inside the conda environment:

pip install -e .

Examples on synthetic data

Run on a synthetic instance, e.g.:

python run.py --nnodes 5 --noise_level 1 --DAG_type path --std --a_size 2 --a_target 3 4 --acquisition greedy

Source code folder: ./optint/

More examples given in: ./optint/notebook/test_multigraphs.ipynb

Examples on Perturb-CITE-Seq [1]

Source code folder: ./perturb-CITE-seq

Notebooks for exploratory data analysis: ./perturb-CITE-seq/preprocess

• download and extract data: ./perturb-CITE-seq/preprocess/screen_sanity_checks.ipynb
• process data: ./perturb-CITE-seq/preprocess/process_data.ipynb

Notebook for running the optimal intervention design task: ./perturb-CITE-seq/test.ipynb

Figures in the paper

Illustraive figures: made using mac keynotes

Pointers for nonillustrative figures:

• ./optint/notebook/test_ow.ipynb: Fig. 3, Supplementary Fig. 2
• ./optint/notebook/test_convergence.ipynb: Fig. 4
• ./optint/notebook/test_multigraphs.ipynb: Fig. 5, Supplementary Fig. 4-7
• ./optint/notebook/test_moreacq.ipynb: Supplementary Fig. 8
• ./optint/notebook/test_misspecgraphs.ipynb: Supplementary Fig. 10
• ./perturb-CITE-seq/preprocess/screen_sanity_checks.ipynb: Supplementary Fig. 11, 13, 14A
• ./perturb-CITE-seq/preprocess/process_data.ipynb: Supplementary Fig. 12
• ./perturb-CITE-seq/preprocess/test_linearity.ipynb: Supplementary Fig. 14C
• ./perturb-CITE-seq/test.ipynb: Fig. 6, Supplementary Fig. 15-18