Testing with kernel-based measures of conditional independence

Code for

"Practical Kernel Tests of Conditional Independence"

Roman Pogodin, Antonin Schrab, Yazhe Li, Danica J. Sutherland, Arthur Gretton

https://arxiv.org/abs/2402.13196

Installation

The slurm scripts assume there's a conda environment splitkci. In addition to python 3.9, pytorch >=2.0, this repository requires

pip install fastargs ratinabox>=1.15.3

The main methods are packaged into a splitkci package. To install, clone this repository and run

pip install -e .

Experiments

The code assumes you have $SCRATCH and $HOME folders, and run the slurm scripts from ./slurm_files. It's recommended to run everything on a GPU (the code will use it if it's available). The experiments require at most 24 GB of RAM and a similar amount of VRAM, but the VRAM requirement can be lowered by using a small chunk_size in wild bootstrap p-value computation (e.g. ci_test.test(..., chunk_size=100) in the code in ./experiments).

Plots for all experiments are made in './notebooks'.

RatInABox

Data for these experiments is generated using ./ratinabox_data/run_slurm_rats_gen.sh. See the readme in ./ratinabox_data.

The scripts for these experiments are start_experiments_rats_budget.sh (budget experiments) and start_experiments_rats.sh (all comparisons).

Toy task

The script for this experiment is start_experiments_randn.sh (including the p-value approximation comparisons).

Car insurance data

This code follows https://github.com/felipemaiapolo/cit.git and re-uses some of their code (see comments in the individual files).

The script is run_slurm_cards.sh and can be run as

sbatch --gres=gpu:ampere:1 -c 8 --mem=48G -t 12:00:00 --partition=main,long --output="./logs/car_insurance.out" \
          --export=ALL,TASK='simulated' ./run_slurm_cars.sh

for the simulated null and

          --export=ALL,TASK='true' ./run_slurm_cars.sh

for the true p-values.

The script will download the dataset from https://github.com/felipemaiapolo/cit.git.