Year: 2023
Type: dissertation
Abstract: Fairness in software systems aims to provide algorithms that operate in a nondiscriminatory manner, with respect to protected attributes such as gender, race, \nor age. Ensuring fairness is a crucial non-functional property of data-driven Machine Learning systems. Several approaches (i.e., bias mitigation methods) have \nbeen proposed in the literature to reduce bias of Machine Learning systems. However, this often comes hand in hand with performance deterioration. Therefore, this \nthesis addresses trade-offs that practitioners face when debiasing Machine Learning \nsystems. \nAt first, we perform a literature review to investigate the current state of the \nart for debiasing Machine Learning systems. This includes an overview of existing \ndebiasing techniques and how they are evaluated (e.g., how is bias measured). \nAs a second contribution, we propose a benchmarking approach that allows for \nan evaluation and comparison of bias mitigation methods and their trade-offs (i.e., \nhow much performance is sacrificed for improving fairness). \nAfterwards, we propose a debiasing method ourselves, which modifies already \ntrained Machine Learning models, with the goal to improve both, their fairness and \naccuracy. \nMoreover, this thesis addresses the challenge of how to deal with fairness with \nregards to age. This question is answered with an empirical evaluation on real-world \ndatasets.
Author: Hort, Max
Institution: University College London
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: green
OpenAlex ID: W7048173825
