Conditional Independence Testing with Heteroskedastic Data and Applications to Causal Discovery
Authors: Wiebke Günther, Urmi Ninad, Jonas Wahl, Jakob Runge
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical causal discovery experiments demonstrate that the adapted partial correlation CI test outperforms the standard test in the presence of heteroskedasticity and is on par for the homoskedastic case.In the following, we conduct experiments evaluating our proposed CI test separately and in conjunction with the PC algorithm. |
| Researcher Affiliation | Collaboration | Wiebke Günther German Aerospace Center Institute of Data Science 07745 Jena, Germany wiebke.guenther@dlr.deUrmi Ninad Technische Universität Berlin 10623 Berlin, Germany urmi.ninad@tu-berlin.deJonas Wahl Technische Universität Berlin 10623 Berlin, Germany wahl@tu-berlin.deJakob Runge German Aerospace Center Institute of Data Science 07745 Jena, Germany and Technische Universität Berlin 10623 Berlin, Germany jakob.runge@dlr.de |
| Pseudocode | Yes | Algorithm 1: Par Corr-WLS |
| Open Source Code | Yes | The code and instructions on how to execute it are provided in the supplemental material. The code includes a method for generating the synthetic data. |
| Open Datasets | No | The paper states 'We generate the data from the SCM (2)' for its experiments and does not provide access information or formal citations for a publicly available or open dataset. |
| Dataset Splits | No | The paper describes generating synthetic data with a 'sample size of 500' and conducting '100 realizations of the SCM' but does not explicitly provide details about training, validation, or test dataset splits or specific data partitioning methodology. |
| Hardware Specification | Yes | The PC algorithm with Par Corr-WLS has an average runtime of 0.29 seconds on homoskedastic data compared to 0.14 seconds for Par Corr-OLS evaluated on AMD 7763. |
| Software Dependencies | No | The paper mentions using the 'Tigramite software package' for their PC-stable algorithm implementation but does not specify a version number for Tigramite or any other key software dependencies. |
| Experiment Setup | Yes | We use the Kolmogorov-Smirnov (KS) statistic to quantify how uniform the distribution of p-values is, and therefore as a metric for type-I errors, as in Runge [2018]. Type-II errors are measured by the area under the power curve (AUPC). The metrics were evaluated with a sample size of 500 from 100 realizations of the SCM. The graph has 10 nodes and 10 edges, a sample size of 500 is used. The significance level α is set to 0.05. The experiment is repeated 500 times. Throughout the experiments this percentage is set to 0.3 to reduce the chance of parent and child being affected by the same kind of heteroskedasticity. For these affected nodes, we choose as a heteroskedasticity type linear or periodic with equal probability (Eq. (3)) and let the noise variance h either depend on one randomly selected parent or the sampling index. We also set a fixed strength s per experiment... All linear dependencies have a coefficient c = 0.5. |