jannisteunissen / mutual_information

Python routines to compute conditional mutual information

Mutual information

Python routines to compute (conditional) mutual information

Overview

Both mutual information $I(x;y)$ and conditional mutual information $I(x;y|z)$ can be computed with this module, using a nearest-neighbor algorithm.

Requirements

• numpy
• scipy (only scipy.special.digamma)
• scikit-learn (only sklearn.neighbors.KDTree)

Usage

Compute mutual information between x and y, which can be 1d or 2d arrays:

mi = compute_mi(x, y, n_neighbors=3, noise_type=None)

Similarly, to compute conditional mutual information:

cmi = compute_cmi(x, y, n_neighbors=3, noise_type=None)

Tests

The file run_tests.py contains several test cases for which analytic solutions are known. Execute it with a -h flag to see available options.

Method

A nearest neighbor approach is used, for which theoretical background is provided in the following papers and references therein:

• Estimating mutual information, by Kraskov et al.
• Estimating Mutual Information for Discrete-Continuous Mixtures, by Gao et al.
• Conditional Mutual Information Estimation for Mixed Discrete and Continuous Variables with Nearest Neighbors, by Mesner and Shalizi

The basic idea is to first determine the Chebyshev distance $\rho_i$ to the k-th nearest neighbor of each sample. For each sample, the following quantity can then be computed:

$\xi_i = \psi(k) + \psi(N) - \psi(n_x + 1) - \psi(n_y + 1)$

where $N$ is the total number of samples, and $n_x$ and $n_y$ are the number of samples within a distance $\rho_i$ if only the $x$ and $y$ coordinates are considered, respectively. The mutual information between is then estimated as the mean value of $\xi_i$

$I(x;y) = <\xi_i>$

Conditional mutual information can be estimated in two ways, either by using the identity

$I(x;y|z) = I(x;y,z) - I(x;z)$

or by using a slightly modified estimator. The radius $\rho_i$ is first computed on the $x, y, z$ data, and then $\kappa_i$ for each sample point is computed as

$\kappa_i = \psi(k) + \psi(n_z+1) - \psi(n_{xz} + 1) - \psi(n_{yz} + 1)$

after which the conditional mutual information of $x$ and $y$ given $z$ is estimated as

$I(x;y|z) = <\kappa_i>$

Implementation

The implementation is inspired by and based on the mutual information methods available in scikit-learn, which where implemented by Nikolay Mayorov. The nearest neighbor searches are performed using the k-d tree implementation provided by scikit-learn.

Acknowledgements

This project has received funding from the European Union’s Horizon 2020 Research and Innovation programme under grant agreement No 776262 AIDA.

Related projects

• https://github.com/sudiptodip15/CCMI
• https://github.com/majianthu/pycopent
• https://github.com/JuliaDynamics/TransferEntropy.jl
• https://github.com/omesner/knncmi
• https://github.com/yandex/CMICOT