Inferring Cause and Effect in the Presence of Heteroscedastic Noise

Authors: Sascha Xu, Osman A Mian, Alexander Marx, Jilles Vreeken

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide a thorough empirical evaluation on synthetic and real world data, comparing to a wide range of methods for bi-variate causal inference. These results show that our method, HECI, performs as well as the strongest competitor whenever noise is homoscedasctic, and is the strongest whenever noise does depend on the cause. In this section, we empirically evaluate HECI on both synthetic data and the real-world T ubingen cause and effect pairs (Mooij et al., 2016) benchmark.
Researcher Affiliation Academia 1CISPA Helmholtz Center for Information Security, Saarbr ucken, Germany 2ETH Z urich & ETH AI Center, Z urich, Switzerland.
Pseudocode Yes Algorithm 1: HECI(X, Y, )
Open Source Code Yes HECI is implemented in Python and we provide the source code as well as the synthetic data for research purposes.1 https://eda.mmci.uni-saarland.de/heci/
Open Datasets Yes We test HECI on two different settings. First, we generate synthetic data according to our assumed causal model in Eq. (1). Next, we use the synthetic data of Gaussian processes provided by Tagasovska et al. (2020) over different noise settings. Last, we benchmark on the real-world T ubingen Cause Effect pairs dataset.
Dataset Splits No The paper describes how synthetic data is generated and mentions using the Tübingen dataset, but it does not specify explicit train/validation/test dataset splits or cross-validation setups for model training and evaluation.
Hardware Specification Yes All experiments were executed on a 4-core Intel i7 machine with 16 GB RAM, running Windows 10.
Software Dependencies No The paper states
Experiment Setup Yes We initiate the binning algorithm with b equal-width bins that partition the domain of X. A local function is fitted inside a single bin or over multiple, neighboring bins. In our experiments, we set = 0.05, with which the best performance was achieved. We therefore require a min support of 10 unique data points per bin. To choose the polynomial degree, we use BIC and minimize