camilbancioiu / mbff

MBFF

This is an experimental framework for studying Markov Blanket Feature Filters. It contains an original implementation of an old, yet canonical feature selection algorithm, based on Markov blankets: Koller and Sahami's algorithm (KS). The trivial Information Gain Thresholding algorithm is implemented as well, as a baseline, even if it doesn't use Markov blankets.

This implementation provides a few efficiency optimizations of the KS algorithm. They can be enabled or disabled at will. The framework also defines a specific set of experiments to evaluate the efficiency gains of these optimizations.

Please submit any comments, improvement suggestions and bugs as Issues of this repository.

Documentation

Will follow soon.

Requirements

1. The framework is written in Python 3. Make sure you have it installed.
2. The Python 3 module sklearn is required. Please install it according to the instructions here: https://scikit-learn.org/stable/install.html.
3. The LYRL2004 version of the Reuters Corpus Volume 1 is required, which will be downloaded by the prepare script, bundled with the framework. Please run ./prepare before attempting anything else. A download of approximately 150 MB will start.