mehmetoktay2016 / rulecovering

Rule covering for interpretation and boosting

Rule Covering for Interpretation and Boosting

We propose two mathematical programming based algorithms for interpretation and boosting of tree-based ensemble methods. These algorithms are called minimum rule cover (MIRCO) and rule cover boosting (RCBoost). The details of both algorithms are given in our paper. In this note, we introduce our implementation of both algorithms as well as list the steps to reproduce our results.

Required packages

All our codes are implemented in Python 3.7 and we use the following packages:

1. scikit-learn
2. numpy
3. gurobipy

We have used the standard installation of Anaconda Distribution (Pyhton 3.7), with which the first two packages are already bundled. The third package can be separately installed again by the Anaconda package manager. Note that along with the Python package, you also need to install Gurobi Optimizer, which is free for research and educational work.

Tutorials

In order to test MIRCO on a set of test problems, we refer to page MIRCO.html or to the notebook. Likewise, for RCBoost we have prepared another page RCBoost.html and a notebook.

Reproducing our results

We provide two scripts MIRCO_run.py and RCBoost_run.py. Running these scripts should reproduce the results that we have reported in our paper.