sibirbil / RuleDiscovery

Two algorithms based on linear programming to discover classification rules for interpretable learning.

Rule Generation for Classification: Scalability, Interpretability, and Fairness

Adia Lumadjeng, Tabea Röber, M. Hakan Akyüz, and Ş. Ilker Birbil

We introduce a new rule-based optimization method for classification with constraints. The proposed method leverages column generation for linear programming, and hence, is scalable to large datasets. The resulting pricing subproblem is shown to be NP-Hard. We recourse to a decision tree-based heuristic and solve a proxy pricing subproblem for acceleration. The method returns a set of rules along with their optimal weights indicating the importance of each rule for learning. We address interpretability and fairness by assigning cost coefficients to the rules and introducing additional constraints. In particular, we focus on local interpretability and generalize separation criterion in fairness to multiple sensitive attributes and classes. We test the performance of the proposed methodology on a collection of datasets and present a case study to elaborate on its different aspects. The proposed rule-based learning method exhibits a good compromise between local interpretability and fairness on the one side, and accuracy on the other side.

You can find the details in our manuscript.

This notebook illustrates how to use RUX and RUG.

This repository contains all script files necessary to reproduce the results reported in the manuscript. However, our method has recently been implemented in the python package ruleopt.

Installation

1. Install Anaconda Distribution.

2. Create a new environment and install the necessary packages:

conda create -n rulediscovery -c conda-forge numpy pandas scikit-learn cvxpy cvxopt

3. Activate the current environment and install gurobi package in the environment:

conda activate rulediscovery

conda install -c gurobi gurobi

Repo structure

The code contains the following files to reproduce the results of our manuscript:

1. In the jupyter notebook RuleDiscovery.ipynb we demonstrate how to use RUG and RUX in single fold on the ecoli dataset. The code will produce the results of RF, ADA and GB, along with RUX(RF), RUX(ADA), RUX(GB), and RUG.

2. The folder num_exp contains all files for the numerical experiments and the case study reported in the manuscript. Please see the README.md in that directory for more details.

Solvers

Note that the default for solver option is 'gurobi'. To use the Gurobi solver, you need to first install it. The solver is freely available for academic use. Check the related page on Gurobi's website. The current version of our code also supports the open source solver GLPK (set solver='glpk').