Identifiability of Causal Graphs under Non-Additive Conditionally Parametric Causal Models

Authors: Juraj Bodik, Valérie Chavez-Demoulin

JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the empirical properties of our methodology on various datasets, demonstrating state-of-the-art performance across multiple benchmarks. Furthermore, we propose an algorithm for estimating the causal structure from random samples drawn from CPCM.
Researcher Affiliation Academia Juraj Bodik EMAIL HEC Lausanne University of Lausanne CH-1015 Lausanne, Switzerland Valerie Chavez-Demoulin EMAIL HEC Lausanne University of Lausanne CH-1015 Lausanne, Switzerland
Pseudocode Yes Algorithm 1: CPCM(F) bivariate case. Algorithm 2: RESIT-greedy algorithm. Algorithm 3: Sequential approach for the choice between S1 and S2. Algorithm 4: Regression with Subsequent Independence Test (RESIT; Peters et al. (2014)), modified for CPCM(F1, . . . , Fk).
Open Source Code Yes The R code for the presented algorithms, simulations, and application is available at https://github.com/jurobodik/Causal_CPCM.git.
Open Datasets Yes We consider seven benchmark datasets, described below. The first five datasets are taken directly from Tagasovska et al. (2020) and described in Appendix B.3. [...] We demonstrate the advantages of CPCM using a subset of the French MTPL motor insurance dataset (Sarpal, 2025).
Dataset Splits No The paper describes simulation setups with specific sample sizes (e.g., 'simulate 100 pairs with n = 1000 data points each' in Section 6.3, and 'focus on a subset of the first n = 1000 records' in Section 6.5), but does not explicitly detail training, testing, or validation splits for these datasets or for evaluation purposes.
Hardware Specification Yes Runtime was measured on a machine with an Intel Core i5-6300U 2.5 GHz processor and 16 GB of RAM.
Software Dependencies No The paper mentions using R, the mgcv package, bnlearn package, and pcalg package, but does not provide specific version numbers for these software components.
Experiment Setup Yes If the p-value is larger than α = 0.05, mark X1 X2 as plausible. [...] With regard to choice of ρ, we use minus the logarithm of the p-value of the independence test (Genest et al., 2019) and λ = 2.