PyRKHSstats

A Python package implementing a variety of statistical/machine learning methods that rely on kernels (e.g. HSIC for independence testing).

Overview

Independence testing with HSIC (Hilbert-Schmidt Independence Criterion), as introduced in A Kernel Statistical Test of Independence, A. Gretton, K. Fukumizu, C. Hui Teo, L. Song, B. Schölkopf, and A. Smola (NIPS 2007).
Measurement of conditional independence with HSCIC (Hilbert-Schmidt Conditional Independence Criterion), as introduced in A Measure-Theoretic Approach to Kernel Conditional Mean Embeddings, J. Park and K. Muandet (NeurIPS 2020).
The Kernel-based Conditional Independence Test (KCIT), as introduced in Kernel-based Conditional Independence Test and Application in Causal Discovery, K. Zhang, J. Peters, D. Janzing, B. Schölkopf (UAI 2011).
Two-sample testing (also known as homogeneity testing) with the MMD (Maximum Mean Discrepancy), as presented in A Fast, Consistent Kernel Two-Sample Test, A. Gretton, K. Fukumizu, Z. Harchaoui, and B. K. Sriperumbudur (NIPS 2009) and in A Kernel Two-Sample Test, A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola (JMLR, volume 13, 2012).

Resource	Description
HSIC	For independence testing
HSCIC	For the measurement of conditional independence
KCIT	For conditional independence testing
MMD	For two-sample testing

The following table details the implementation schemes for the different resources available in the package.

Resource	Implementation Scheme	Numpy based available	PyTorch based available
HSIC	Resampling (permuting the x_i's but leaving the y_i's unchanged)	Yes	No
HSIC	Gamma approximation	Yes	No
HSCIC	N/A	Yes	Yes
KCIT	Gamma approximation	Yes	No
KCIT	Monte Carlo simulation (weighted sum of χ² random variables)	Yes	No
MMD	Gram matrix spectrum	Yes	No