ListConditionalIndependencies

This repository contains the implementation for the paper "Testing Causal Models with Hidden Variables in Polynomial Delay via Conditional Independencies" by Hyunchai Jeong*, Adiba Ejaz*, Jin Tian, and Elias Bareinboim. More specifically, the algorithms for listing all conditional independence relations (CIs) implied by a causal model. These authors* contributed equally.

How to install

Please run the following commands to install the package:

git clone https://github.com/CausalAILab/ListConditionalIndependencies.git
cd ListConditionalIndependencies
pip3 install -r requirements.txt
pip3 install -e .

How to run a listing algorithm

To run some examples, try running the following command:

python3 main.py clmp graphs/paper/fig4b.txt

Arguments

1. First argument: gmp, lmp, or clmp. Specifies a listing algorithm that outputs all CIs invoked by the following property.

• gmp: The global Markov property (runs the algorithm ListGMP)
• lmp: The ordered local Markov property (runs the algorithm ListCIBF)
• clmp: The c-component local Markov property (runs the algorithm ListCI)

2. Second argument: graphs/paper/fig4b.txt. The path to a text file that contains graph information, such as nodes and edges of a graph. Please check the formatting for details.

Format of the graph file

Consider graphs/paper/fig4b.txt as an example.

<NODES>

Each line represents the name of a variable/node (e.g., A and B).

<EDGES>

Each line describes an edge. Two types of edges are supported:

1. Directed edge: A -> B represents a directed edge from A to B.
2. Bidirected edge: A -- F means there exists some unmeasured confounder (i.e., latent variable) between A and F.

How to test a model against observational data

Try running the following command:

python3 test_model.py graphs/sachs/graph_gt.txt datasets/sachs/dataset.csv

The example tests a ground-truth graph provided by experts (11 nodes and 17 edges) against a real-world protein signaling dataset (853 samples) by (Sachs et al. 2005). The details are provided in Appendix F.2 in the paper.

Arguments

1. First argument: graphs/sachs/graph_gt.txt. The path to a graph file.
2. Second argument: datasets/sachs/dataset.csv. The path to a dataset file.

How to run experiments

We consider experiments shown in Appendix F.1 and F.3.

1. Appendix F.1: Comparison of ListCI and two other algorithms - ListGMP and ListCIBF.

   • Running bnlearn instances (with varying projection levels) over three algorithms. The original bnlearn instances are available here.

   Try running the following command:

   python3 run_bnlearn_graph.py clmp graphs/bnlearn/sm/asia.txt
   

   When the experiment finishes, a report file named bnlearn_report_asia_clmp.csv will be generated. Please check the heading Parameters of a report file for more details.

   Arguments

   1. First argument: gmp, lmp, or clmp. An algorithm to run.
   2. Second argument: graphs/bnlearn/sm/asia.txt. The path to a bnlearn graph file.

   Graphs

   All bnlearn graphs are placed in the directory graphs/bnlearn. There are three types of graphs, classified based on n: number of nodes.

   • Small graphs, n: [1,20)

     Placed in graphs/bnlearn/sm subdirectory.

   • Medium graphs, n: [21,50)

     Placed in graphs/bnlearn/md subdirectory.

   • Large graphs, n: [51,100)

     Placed in graphs/bnlearn/lg subdirectory.

   Parameters of a report file

   Measured parameters for all three algorithms.

   • n: number of nodes.
   • m: number of edges.
   • md: number of directed edges.
   • mu: number of bidirected edges.
   • # CI: number of CIs invoked by a Markov property.
   • runtime: running time of an algorithm in seconds (rounded to the nearest second).

   Two additional paramters are measured for ListCIBF.

   • # S: number of ancestral sets.
   • # S+: number of maximal ancestral sets (MASs).

   One additional paramter is measured for both ListCIBF and ListCI.

   • s: size of the largest c-component.

   Experiment configuration

   The following parameter may be modified in run_bnlearn_graph.py to change experiment settings.

   • numBatches: number of batches to run for an experiment. For each single batch, 10 random graphs are generated in total: one graph for each projection level U (for U = 0, 10, ..., 90). Default is set to 10, totalling 100 sample graphs. Located at line 355 of run_bnlearn_graph.py.
   • fixOrdering: If set to True, a topological order over variables will be fixed (across all projections) for more consistent results. Default is True.
   • randomSeed: A random seed to be provided to random.random(). Default is None.

2. Appendix F.3: Analysis of C-LMP

   • Running experiments over random graphs to understand the total number of valid CIs invoked by C-LMP.

   Try running the following command:

   python3 run_case.py 1a
   

   A report file named report_case_1a.csv will be generated.

   Arguments

   1. First argument: Any one string from the list ['1a', '1b', '1v', '1r', '2a', '2b', '3a'].
   • Case 1
     • 1a: Experiments corresponding to plots shown in Fig. F.3.1.
     • 1v, 1r: Fig. F.3.2a and Fig. F.3.2b, respectively.
     • 1b: Fig. F.3.3.
   • Case 2
     • 2a: Fig. F.3.4.
     • 2b: Fig. F.3.5.
   • Case 3
     • 3a: Fig. F.3.6.

   Experiment configuration

   The following parameters may be modified to change experiment settings. The mentioned variables are defined below the line: if __name__ == '__main__':.

   • n: number of nodes.
   • numBatches: number of batches to run for an experiment. For each single batch, 10 random graphs are generated. Default is set to 10, totalling 100 sample graphs.
   • randomSeed: A random seed to be provided to random.random(). Default is None.

References

• Jeong, H., Ejaz, A., Tian, J., and Bareinboim, E. 2024. Testing Causal Models with Hidden Variables in Polynomial Delay via Conditional Independencies. Causal Artificial Intelligence Lab, Columbia University.
• Scutari, M. 2010. Learning Bayesian Networks with the bnlearn R Package. Journal of Statistical Software, 35(3): 1–22.
• Sachs, K., Perez, O., Pe’er, D., Lauffenburger, D. A., and Nolan, G. P. 2005. Causal protein-signaling networks derived from multiparameter single-cell data. Science, 308(5721): 523–529.