Fast Proxy Experiment Design for Causal Effect Identification

Authors: Sepehr Elahi, Sina Akbari, Jalal Etesami, Negar Kiyavash, Patrick Thiran

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present numerical experiments that showcase the empirical performance and time efficiency of our proposed exact and heuristic algorithms.
Researcher Affiliation Academia Sepehr Elahi EPFL, Switzerland sepehr.elahi@epfl.ch Sina Akbari EPFL, Switzerland sina.akbari@epfl.ch Jalal Etesami TUM, Germany j.etesami@tum.de Negar Kiyavash EPFL, Switzerland negar.kiyavash@epfl.ch Patrick Thiran EPFL, Switzerland patrick.thiran@epfl.ch
Pseudocode Yes Algorithm 1 Intervention design for generalized adjustment
Open Source Code No We will release our code as a toolbox on Git Hub, providing access for researchers and practitioners. The link will be available upon publication of the paper.
Open Datasets Yes We ran each algorithm for solving the MCID problem on 100 randomly generated Erdos-Renyi [Erdos and Renyi, 1960] ADMG graphs with directed and bidirected edge probabilities ranging from 0.01 to 1.00, in increments of 0.01. We conduct experiments using 17 real-world networks from the Bayesian Network Repository8. This repository encompasses networks from diverse domains such as biology, engineering, medicine, and social science. Footnote 8: bnlearn.com/bnrepository
Dataset Splits No The paper does not provide specific percentages or counts for training, validation, or test splits. It mentions 'synthetic and real-world experiments' but lacks details on data partitioning.
Hardware Specification Yes All experiments, coded in Python, were conducted on a machine equipped two Intel Xeon E5-2680 v3 CPUs, 256GB of RAM, and running Ubuntu 20.04.3 LTS.
Software Dependencies No Our codebase is implemented fully in Python. We use the Py SAT library for formulating and solving the WPMAX-SAT problem, and the Pu LP library for formulating and solving the ILP problem. We used the RC2 algorithm [Ignatiev et al., 2019], and the Gurobi solver [Gurobi Optimization, LLC, 2023], to solve the WPMAX-SAT problem, and the ILP, respectively.
Experiment Setup Yes We ran each algorithm for solving the MCID problem on 100 randomly generated Erdos-Renyi [Erdos and Renyi, 1960] ADMG graphs with directed and bidirected edge probabilities ranging from 0.01 to 1.00, in increments of 0.01. We performed two sets of simulations: for single-district and multiple-district settings, respectively. In the single-district case, we varied n, the number of vertices, from 20 to 100, while in the multiple-district case, we fixed n = 20 and varied the number of districts from 1 to 9.