lgbn 0.1.2

Structure and parameter learning for linear Gaussian Bayesian networks.

lgbn

Structure and parameter learning for linear Gaussian Bayesian networks.

This package provides structure learning of linear Gaussian Bayesian networks via score-based search algorithms. Available algorithms include K2, Greedy Hill Climbing and Greedy Equivalent Search. Available scores include the Log Likelihood score and the Bayesian Information Criterion score.

This is a well-tested (with 100% coverage) and well-documented (full API documentation + design and usage guides) package that focuses on training for a particular kind of Bayesian network: those in which every node has a Gaussian distribution where the mean is a linear combination of the parent nodes plus a bias factor and where the variances are independent across all nodes. This particular kind of Bayesian network has applications in all sorts of problems involving continuous data, even if the distribution of the data itself is not Gaussian.

To our knowledge, this is the only available Python package able to work with these kinds of structures and data. Other Bayesian network packages such as pgmpy have not yet implemented continuous Gaussian nodes with the restrictions mentioned before. There exists software in other programming languages such as the Bayesian Network Toolbox (BNT) for MatLab, last released in 2003 or the R package bnlearn which is actively maintained. We have used both as reference implementations and verified that the results obtained with this package match those returned by BNT and bnlearn with a relative error less than one part per million.

Documentation

Documentation for this package is available online at https://lgbn.readthedocs.io/en/latest/

Development

Development has been done on a Linux machine using Python 3.9. Although we strived to make the code portable, the package might not work fully on other platforms. This section covers how to get up and running contributing to this package on a Linux-compatible setup (MacOS and WSL should mostly work).

Development setup

Clone the repository and run the following commands inside the repository root:

python3 -m pip install virtualenv
python3 -m virtualenv .venv
source .venv/bin/activate

pip install -r requirements.txt

These commands will create a virtual environment and install all dependencies isolated from the rest of your system. Verify everything is working by running the test suite (see next section).

Testing

This repository includes unit tests for the Python package written using the unittest module of the Python standard library. To run the tests we rely on unittest's autodiscovery feature:

python3 -m unittest

Test coverage can be measured with coverage.py:

coverage run -m unittest discover
coverage report -m

Building the package

Ensure you have installed all dev dependencies. Then run

python3 -m build

from the repository root. This will build the package and place it under dist/lgbn-x.y.z.tar.gz. You can install it on your own machine by running

python3 -m pip install --user dist/lgbn-x.y.z.tar.gz

Building documentation

Documentation is writen in reStructuredText and consists of two parts:

• API reference documentation is written in the docstrings of classes, methods, etc. following the numpy style guide. These are then imported into the documentation using autodoc.

• Usage and design documentation, consisting of long form documents which explain how to work with the package and the design decisions behind it. These documents are stored in the docs/ directory.

Documentation can be compiled in to a set of HTML documents plus many other formats by using Sphinx. To build the HTML version documentation use

cd docs
make html

Uploading the package to PyPI

Build the package and execute

python3 -m twine upload dist/*

After that you will be asked for credentials and the updated package will be available for installation with pip install lgbn.

Acknowledgements

This package was developed while writing the bachelor's thesis of the author at Universidad Autónoma de Madrid. The Bachelor's thesis was supervised by Daniel Ramos with help from Pablo Ramírez Hereza, both members of the Audio, Data Intelligence and Speech research group (AUDIAS) at Escuela Politécnica Superior (UAM).

Motivation came from the lack of support for linear Gaussian Bayesian networks in other Python packages like pgmpy, which has served as a source of inspiration for the structure of the package.

License

This package and its documentation are copyright of Elias Hernandis. They are released under the MIT license. The MIT License grants anyone permission to use, copy, modify and distribute versions of this software without limitations, including for comercial purposes. This software is provided without any warranties. Please see the LICENSE file for more information.