Vector Causal Inference between Two Groups of Variables

Authors: Jonas Wahl, Urmi Ninad, Jakob Runge

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our methods empirically and compare them to other state-of-the-art techniques.In Section 5, we analyse the empirical performance of these algorithms in experiments with synthetic data and compare it to that of other approaches (Vanilla-PC and the Trace Method (Janzing et al. 2009; Zscheischler, Janzing, and Zhang 2012)). We also consider a real world climate science example of surface temperatures in the El Ni no Southern Oscillation (ENSO 3.4) region in the pacific and in British Columbia to test our algorithms.
Researcher Affiliation Academia Jonas Wahl*,1,2, Urmi Ninad*,1,2, Jakob Runge1,2 1Technische Universit at Berlin 2 DLR Institut f ur Datenwissenschaften Jena
Pseudocode Yes Algorithm 1: 2G-Vec CI.PC
Open Source Code Yes All code is available at https://github.com/Jonas Choice/ 2GVec CI.
Open Datasets Yes NCEP-NCAR Reanalysis 1 data was provided by NOAA PSL, Boulder, Colorado, USA, from their website at https://psl.noaa.gov, see Kalnay et al. (1996).
Dataset Splits No The paper describes the generation of simulated data and the use of real-world data with varying sample sizes, but it does not specify explicit train/validation/test dataset splits, nor does it mention cross-validation or other detailed splitting methodologies for reproducibility.
Hardware Specification Yes Computations were done on Bull Sequana XH2000 with AMD 7763 CPUs.
Software Dependencies No The paper mentions using specific tests and algorithms (e.g., 'partial correlation test', 'Gaussian Process distance correlation independence test', 'PC-algorithm') but does not specify version numbers for any software libraries, frameworks, or dependencies used.
Experiment Setup Yes Models vary along the following parameters: sample size (between 50 and 500); group sizes n and m (between 3 and 100); edge densities within X and ηY (between 1% and 90% of all possible edges); density of the interaction matrix A (between 1% and 90% of all possible entries non-zero); effect size, i.e. size of the entries in A (uniformly randomly drawn from different intervals).In both algorithms, we test for conditional independencies using the partial correlation test at significance level α = 0.01. We choose the sensitivity parameter to be α = 0.01 if not specified differently.