Fairness of Machine Learning Algorithms in Demography

TL;DR

This study improves fairness in demographic machine learning models by reducing reliance on sensitive features through feature dropout, maintaining or enhancing accuracy, and validating on Russian demographic data.

Contribution

Introduces a feature dropout method inspired by neural dropout to enhance fairness in classifiers while preserving accuracy, evaluated with LIME explanations on real demographic data.

Findings

Classifiers became less dependent on sensitive features.

Fairness improved without sacrificing accuracy.

Method effective across multiple classifier types.

Abstract

The paper is devoted to the study of the model fairness and process fairness of the Russian demographic dataset by making predictions of divorce of the 1st marriage, religiosity, 1st employment and completion of education. Our goal was to make classifiers more equitable by reducing their reliance on sensitive features while increasing or at least maintaining their accuracy. We took inspiration from "dropout" techniques in neural-based approaches and suggested a model that uses "feature drop-out" to address process fairness. To evaluate a classifier's fairness and decide the sensitive features to eliminate, we used "LIME Explanations". This results in a pool of classifiers due to feature dropout whose ensemble has been shown to be less reliant on sensitive features and to have improved or no effect on accuracy. Our empirical study was performed on four families of classifiers (Logistic Regression, Random Forest, Bagging, and Adaboost) and carried out on real-life dataset (Russian demographic data derived from Generations and Gender Survey), and it showed that all of the models became less dependent on sensitive features (such as gender, breakup of the 1st partnership, 1st partnership, etc.) and showed improvements or no impact on accuracy