Front-door Adjustment Beyond Markov Equivalence with Limited Graph Knowledge

Authors: Abhin Shah, Karthikeyan Shanmugam, Murat Kocaoglu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach empirically in 3 ways: (i) we demonstrate the applicability of our method on a class of random graphs, (ii) we assess the effectiveness of our method in estimating the ATE using finite samples, and (iii) we showcase the potential of our method for causal fairness analysis.
Researcher Affiliation Collaboration Abhin Shah Massachusetts Institute of Technology abhin@mit.edu Karthikeyan Shanmugam Google Research karthikeyanvs@google.com Murat Kocaoglu Purdue University mkocaoglu@purdue.edu
Pseudocode Yes Algorithm 1: ATE estimation using subset search. Input: nr, t, y, b, Z, pv Output: ATEz, ATEs
Open Source Code Yes The source code of our implementation is available at https://github.com/ abhin-shah/FD-adjustment-with-limited-graph.
Open Datasets Yes The German Credit dataset [Hofmann, 1994] is used for credit risk analysis... We perform a similar analysis on the Adult dataset [Kohavi and Becker, 1996].
Dataset Splits No The paper mentions 'a specific train-test split' in Algorithm 1 and 'half of the training data' for bootstrapping in German Credit dataset analysis, but it does not provide explicit percentages, sample counts, or references to predefined standard train/validation/test splits for the datasets used in the experiments.
Hardware Specification Yes In this work, we used a workstation with an AMD Ryzen Threadripper 3990X 64-Core Processor (128 threads in total) with 256 GB RAM and 2x Nvidia RTX 3090 GPUs. However, our simulations only used the CPU resources of the workstation.
Software Dependencies No The paper lists several software dependencies (networkx, causal-learn, RCoT, ridge CV) but does not provide explicit version numbers for these components, which is required for reproducibility according to the criteria.
Experiment Setup Yes The paper provides details on synthetic data generation, including the use of Unif[1,2] for unobserved variables, linear combinations with coefficients from Unif[1,2], and Gaussian noise for observed variables (Eq. 11). It also specifies hyperparameters like 'pv = 0.1' for the German Credit dataset analysis and the number of runs 'nr = 100'.