DGCIT: Double Generative Adversarial Networks for Conditional Independence Testing

This repository contains an implementation and further details of Double Generative Adversarial Networks for Conditional Independence Testing.

Reference: Shi, C., Xu, T., Bergsma, W. and Li, L. (2021+) Double Generative Adversarial Networks for Conditional Independence Testing. Journal of Machine Learning Research, accepted.

Paper link: https://arxiv.org/pdf/2006.02615.pdf

Setup

$ git clone https://github.com/tianlinxu312/dgcit.git
$ cd dgcit/

# Create a virtual environment called 'venv'
$ virtualenv venv 
$ source venv/bin/activate    # Activate virtual environment

# Install all dependencies
$ python3 -m pip install -r requirements.txt 

Data

CCLE data used in the paper is downloaded from here: https://github.com/alexisbellot/GCIT/tree/master/CCLE%20Experiments

Training

For runing experiments to compute Type I error:

# Compute Type I error for 1000 samples
$ python3 train.py \
    --model="dgcit"
    --test="type1error"
    --n_samples=1000

For runing experiments to compute Power:

# Compute Power for 1000 samples
$ python3 train.py \
    --model="dgcit"
    --test="power"
    --n_samples=1000

For more baseline models and parameter settings, please see train.py file.

Illustration of conditional independence testing with double GANs

Type I error and power

Top panels: the empirical type-I error rate of various tests under H0. From left to right: normal Z with α = 0.1, normal Z with α = 0.05, Laplacian Z with α = 0.1, and Laplacian Z with α = 0.05. Bottom panels: the empirical power of various tests under H1.

KCIT results were obtained using this implementation: https://github.com/ericstrobl/RCIT/blob/master/R/KCIT.R

Results for the anti-cancer drug example

The variable importance measures of the elastic net(EN) and random forest(RF) models, versus the p-values of the GCIT and DGCIT tests for the anti-cancer drug example.