Linear Causal Representation Learning from Unknown Multi-node Interventions

Authors: Burak Varıcı, Emre Acartürk, Karthikeyan Shanmugam, Ali Tajer

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

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Simulations We empirically assess the performance of the UMNI-CRL algorithm for recovering the latent DAG G and latent variables Z.
Researcher Affiliation Collaboration Burak Varıcı Carnegie Mellon University Emre Acartürk Rensselaer Polytechnic Institute Karthikeyan Shanmugam Google Deep Mind Ali Tajer Rensselaer Polytechnic Institute
Pseudocode Yes Algorithm 1 Unknown Multi-node Interventional (UMNI)-CRL
Open Source Code Yes The codebase for the experiments can be found at https://github.com/acarturk-e/umni-crl.
Open Datasets No The paper describes generating its own synthetic data for experiments ('To generate G, we use Erd os-Rényi model...', 'observed variables are generated as X = G Z'), and does not mention using a publicly available or open dataset with a specific link, DOI, or formal citation.
Dataset Splits No The paper states 'generate ns = 105 samples of Z from each environment' but does not specify any training, validation, or test dataset splits.
Hardware Specification No The paper states 'Experiments are run on a single commercial CPU' but does not provide specific hardware details such as exact CPU models, processor types with speeds, or memory amounts.
Software Dependencies No The paper mentions statistical models and general types of estimators for score functions but does not provide specific software names with version numbers, such as library or solver names.
Experiment Setup Yes For the causal models, we adopt linear structural equation models (SEMs) with Gaussian noise. The nonzero edge weights of the linear SEMs are sampled from Unif( [0.5, 1.5]), and the noise terms are zero-mean Gaussian variables with variances σ2 i sampled from Unif([0.5, 1.5])... We consider target dimensions d {10, 50}, generate ns = 105 samples of Z from each environment... we use a partial correlation test and set the significance level to α = 0.05... we set κ = 2 in the simulations in Section 5.