Fair Feature Subset Selection using Multiobjective Genetic Algorithm
This work addresses fairness in feature selection for machine learning decision-making, which is an incremental improvement by integrating fairness metrics into existing optimization methods.
The paper tackles the problem of selecting feature subsets to improve both fairness and accuracy in machine learning models, using a multiobjective genetic algorithm to find Pareto-optimal solutions that balance statistical disparity and F1-Score, with experiments on benchmark datasets showing effective exploration of this trade-off.
The feature subset selection problem aims at selecting the relevant subset of features to improve the performance of a Machine Learning (ML) algorithm on training data. Some features in data can be inherently noisy, costly to compute, improperly scaled, or correlated to other features, and they can adversely affect the accuracy, cost, and complexity of the induced algorithm. The goal of traditional feature selection approaches has been to remove such irrelevant features. In recent years ML is making a noticeable impact on the decision-making processes of our everyday lives. We want to ensure that these decisions do not reflect biased behavior towards certain groups or individuals based on protected attributes such as age, sex, or race. In this paper, we present a feature subset selection approach that improves both fairness and accuracy objectives and computes Pareto-optimal solutions using the NSGA-II algorithm. We use statistical disparity as a fairness metric and F1-Score as a metric for model performance. Our experiments on the most commonly used fairness benchmark datasets with three different machine learning algorithms show that using the evolutionary algorithm we can effectively explore the trade-off between fairness and accuracy.