Local Discovery by Partitioning: Polynomial-Time Causal Discovery around Exposure-Outcome Pairs

Files

• ldp.py: The main function of class LDP is partition_z().
• ldp_utils.py: Class LDPUtils provides some basic helper functions for experiments.
• data_generation.py: Class DataGeneration provides a basic linear-Gaussian data generating process for demonstration purposes.
• ldp_demo.py: This script provides a demo of LDP functionality on a linear-Gaussian DAG.
• Software environment: We provide both environment.yml and requirements.txt, either of which can be used to recreate the environment used to execute LDP and reproduce all experiments.

Demo

We provide a script to demo LDP on a linear-Gaussian DAG using the Fisher-z independence test. This DAG can optionally contain an M-structure, a butterfly structure, or both. X can be a direct cause of Y, or have no direct effect.

python ldp_demo.py -x=1 -m=0 -b=0 -n=5000 -a=0.005 -r=10 -e=0

Arguments:

• -x (int): whether X directly causes Y or not (1 = True, 0 = False).
• -m (int): whether the DAG contains an M-structure or not (1 = True, 0 = False).
• -b (int): whether the DAG contains a butterfly structure or not (1 = True, 0 = False).
• -n (int): sample size.
• -a (float): alpha for p-value of independence test.
• -r (int): total replicate DAGs to run.
• -e (int): whether to export results or not (1 = True, 0 = False).