Skip to content

Latest commit

 

History

History
150 lines (101 loc) · 4.88 KB

README.md

File metadata and controls

150 lines (101 loc) · 4.88 KB

Causal Structure Learning for SAGE Estimation

Code accompanying the paper 'Efficient SAGE Estimation via Causal Structure Learning'

Summary

R code to generate data from R environments (.RDA files) downloaded from the bnlearn repository as well as completely synthetic data. R files to infer graphs from (semi-)synthetic as well as real-world data. Benchmark of graph learning with respect to learned d-separations. Experiment files to fit models and infer SAGE values for the features of the respective models. Experiment file to exploit d-separations/independeces in SAGE inference. File to test the spared time of latter approach as opposed to standard SAGE estimation. For all experiments conducted using Python we used version 3.9.12.

Replicate

To replicate the study, clone this repository, so that all directories adhere the requirements of the different R and Python scripts.

git clone [link]

To make use of the RFI package, clone it into this the directory of this (cloned) repository and install.

git clone [link]

and

pip install -e rfi

where rfi is the path to the directory of the cloned RFI repository. Now proceed as explained below:

Data

We used the software R (version 4.1.2) to generate the data (if necessary) and conduct the causal structure learning.

To generate semi-synthetic categorical data from .RDA files (located in ~/datagen/envs/) execute (not used in paper):

RScript datagen/datagen_env.R

To generate synthetic linear Gaussian data using the R package 'pcalg' (https://cran.r-project.org/web/packages/pcalg/index.html) execute:

bash datagen_cont.sh

The real-world datasets are available from: https://archive.ics.uci.edu/ml/datasets/Drug+consumption+%28quantified%29#

Causal Structure Learning

To conduct causal structure learning, we again relied on the software R (version 4.10). To execute the respective learning algorithm, choose the file accordingly. For semi-synthetic categorical data, the files are named by the respective algorithm only (e.g. tabu.R for a greedy search algorithm using TABU search heuristic). For the synthetic continuous data, the R files have the addendum '_cont' after the algorithm name (e.g. tabu_cont.R). The application of the algorithms to the employed real-world data is done using the files with the addendum '_real' (e.g. tabu_real.R). Each file covers all data files that pertain to the respective descriptions and each .csv file has to exist and be in the corresponding directory according to the structure in this repository.

Application of tabu search algorithm to every file containing synthetic linear Gaussian data via bash script:

bash bnlearn.sh

and for the real-world application:

Rscript bnlearn/tabu_drug.R

Evaluation of Causal Structure Learning

Evaluation of causal structure learning is done with respect to d-separations between a set of potential explanatory variables and a dedicated target. Since targets are (mostly) sampled at random for the (semi-)synthetic data, you need to execute the sampling of targets first (semi-synthetic data not used in Efficient SAGE Estimation via Causal Structure Learning):

python sample_targets.py

Then you can execute:

python graph_evaluation.py

SAGE Inference

Use respective experiment files, sage_cont.py for synthetic continuous data and sage_real.py for real-world data. We use the SAGE implementation from (link). The execution for synthetic continuous data can be done using the following bash script.

bash sage_experiments.sh

Now convert the results (based on intermediate results from the SAGE inference) to csl-sage/experimental results. This is done by setting inferred summands of the SAGE representation:

(equation)

to zero given X_j dsep Y given coalition. This is done post-hoc to the SAGE results in order to save runtime during the experiments but also implemented in the original rfi package to be used in practice. To this end, execute:

bash sage_to_dsage.sh

The file ai_via_timing.py serves the purpose to evaluate which part of the SAGE approximation that is potentially skipped when a d-separation can be found. To replicate the results, execute

runtime_estimation.sh

Visualization

Successively execute the following files from visualization folder with the according command line arguments.

python visualization/bnlearn/confusion_horizontal.py --alg tabu --size 10000
python visualization/bnlearn/runtime_and_f1.py --alg tabu
python visualization/sage/runtime_sage.py 
python visualization/sage/sage_values.py --model lm --degree 2 --top 5 
python visualization/sage/deltas_boxplot.py --model lm --degree 2

Real-world example

Successively execute the following files from real-world-experiments folder.

python real-world-experiments/experiment_drug.py
python sage_to_csl_drug.py
sage_dsage_plot.py
deltas_drugs.py