Causal discovery with endogenous context variables

Authors: Wiebke Günther, Oana-Iuliana Popescu, Martin Rabel, Urmi Ninad, Andreas Gerhardus, Jakob Runge

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments demonstrate the performance of the proposed method over alternative baselines, but they also unveil current limitations of our method. [...] 5 Simulation study
Researcher Affiliation Collaboration 1German Aerospace Center (DLR), Institute of Data Science, Jena, Germany 2Technische Universität Berlin, Institute of Computer Engineering and Microelectronics, Berlin, Germany 3Center for Scalable Data Analytics and Artificial Intelligence (Sca DS.AI) Dresden / Leipzig, Germany 4Technische Universität Dresden, Faculty of Computer Science, Dresden, Germany
Pseudocode Yes Algorithm 1 Adaptive CD for Discovering Context-Specific Graphs with Endogenous Context Variables
Open Source Code Yes The code, based on causal-learn (40) and Tigramite (25), and details for replication (random seeds) can be found here: https://github.com/oanaipopescu/adaptive_endogenous_contexts.
Open Datasets No Our data models are created by first randomly generating a linear acyclic SCM with up to D + 1 nodes (base graph) of the form Xi = PD+1 j aij Xj + cjηj, where i = 1, . . . , D + 1, at a desired sparsity level s. [...] Once a context-specific SCM is created, we generate the data according to this SCM. The paper describes a process for generating synthetic data for each experiment, rather than using a fixed, publicly available dataset.
Dataset Splits No The paper focuses on generating synthetic data for its experiments and does not specify any fixed training, validation, or test dataset splits. Performance is evaluated based on metrics over the generated samples.
Hardware Specification Yes All results were obtained by conducting trials in parallel across a cluster of 116 CPU nodes, each equipped with 2x Intel Xeon Platinum 8260 processors.
Software Dependencies No The code, based on causal-learn (40) and Tigramite (25), and details for replication (random seeds) can be found here: https://github.com/oanaipopescu/adaptive_endogenous_contexts. The specific version numbers for causal-learn and Tigramite, or other software dependencies, are not provided in the paper.
Experiment Setup Yes We present results where D + 1 = 8 and density s = 0.4. For all j, ηj N(0, 1). We draw the non-zero ai,j randomly from {1.8, 1.5, 1.2, 1.2, 1.5, 1.8}, and set cj = 1 for all j. For the context indicator R, we set c R = 0.2.