Fair and Interpretable Decision Rules for Binary Classification

Explore a 31-minute conference talk on fair and interpretable decision rules for binary classification, presented by Oktay Günlük from Cornell University at the Deep Learning and Combinatorial Optimization 2021 event. Delve into the problem of constructing Boolean rule sets in disjunctive normal form (DNF) for interpretable binary classification models, subject to fairness constraints. Learn about the innovative integer programming approach that maximizes classification accuracy while explicitly addressing group fairness. Discover the column generation framework with a novel formulation that efficiently searches through exponentially many possible rules, eliminating the need for heuristic rule mining. Compare this method to other interpretable machine learning algorithms and understand its superior performance in balancing fairness and accuracy. Follow the presentation's structure, covering topics such as interpretation, assumptions, feature space, pricing problem, computational approach, data sets, Hamming loss vs. accuracy, and control of fairness. Gain insights into this collaborative work between Oktay Günlük and Connor Lawless, presented at the Institute for Pure and Applied Mathematics, UCLA.

Intro
Fair and Interpretable Classification
Binary Classification
Interpretation
Assumptions
Feature Space
Integer Program
Pricing Problem
Computational Approach
Data Sets
Humming Loss vs Accuracy
Control of Fairness
Comparison
Conclusion