FlopsKa / nystroem-mhsic

Code for the article "Nyström M-Hilbert-Schmidt Independence Criterion".

General

This repository hosts the code for the article

• Florian Kalinke, Zoltán Szabó. Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, PMLR 216:1005-1015, 2023. Link.

Instructions for Reproducing the Experiments

Data

The data must be obtained from the sources within the article. For the causality experiments, see below.

Causality Data

We generate the data to run the causality experiments with R, using code from [Pfister, 2018], save it as .csv and continue with python. The code is in lib/pfister2018.

Simulated Causality Data

Note that we change line 214 of the original file sampleDataFromG.R to obtain samples with larger variance:

noisetmp <- (1*sqrt(noiseVar[node]) * abs(ran))^(parsNoise$noiseExp) * sign(ran)

Create the folder structure with

cd data
mkdir -p causal_sim/500/{1..100} causal_sim/1000/{1..100} causal_sim/1500/{1..100}

And generate the data (this may take some time) by

cd causality_example
Rscript gen_sim_data.R

Weather Data

The commands to obtain the residuals are

cd data
mkdir causal_real
cd causality_example
Rscript gen_real_data.R

Experiments

The code to reproduce the experiments is in figxx.py, respectively. The plots can be created with plot_figxx.ipynb.

For NFSIC, we use the code from fsic-test with commit id 1bc6318 [Jitkrittum, 2017].

[Pfister, 2018] Niklas Pfister, Peter Bühlmann, Bernhard Schölkopf, and Jonas Peters. Kernel-based tests for joint independence. Journal of the Royal Statistical Society. Series B. Statistical Methodology, pages 5–31, 2018.

[Jitkrittum, 2017] Jitkrittum, Wittawat, Zoltán Szabó, and Arthur Gretton. An adaptive test of independence with analytic kernel embeddings. International Conference on Machine Learning. PMLR, 2017.