Not So Fair: The Impact of Presumably Fair Machine Learning Models

TY - CHAP

T1 - Not So Fair

T2 - The Impact of Presumably Fair Machine Learning Models

AU - Jorgensen, MacKenzie

AU - Richert, Hannah

AU - Black, Elizabeth

AU - Criado, Natalia

AU - Such, Jose

N1 - Funding Information: This work was supported by UK Research and Innovation [grant number EP/S023356/1] in the UKRI Centre for Doctoral Training in Safe and Trusted Artificial Intelligence. The second author was funded by the DAAD RISE Worldwide 2022 program to complete research with the first author over summer 2022 in London. We would like to thank the reviewers for their constructive feedback and Maria Stoica and Julia Barnett for their helpful feedback. Publisher Copyright: © 2023 ACM.

PY - 2023/8/29

Y1 - 2023/8/29

N2 - When bias mitigation methods are applied to make fairer machine learning models in fairness-related classification settings, there is an assumption that the disadvantaged group should be better off than if no mitigation method was applied. However, this is a potentially dangerous assumption because a "fair"model outcome does not automatically imply a positive impact for a disadvantaged individual - they could still be negatively impacted. Modeling and accounting for those impacts is key to ensure that mitigated models are not unintentionally harming individuals; we investigate if mitigated models can still negatively impact disadvantaged individuals and what conditions affect those impacts in a loan repayment example. Our results show that most mitigated models negatively impact disadvantaged group members in comparison to the unmitigated models. The domain-dependent impacts of model outcomes should help drive future bias mitigation method development.

AB - When bias mitigation methods are applied to make fairer machine learning models in fairness-related classification settings, there is an assumption that the disadvantaged group should be better off than if no mitigation method was applied. However, this is a potentially dangerous assumption because a "fair"model outcome does not automatically imply a positive impact for a disadvantaged individual - they could still be negatively impacted. Modeling and accounting for those impacts is key to ensure that mitigated models are not unintentionally harming individuals; we investigate if mitigated models can still negatively impact disadvantaged individuals and what conditions affect those impacts in a loan repayment example. Our results show that most mitigated models negatively impact disadvantaged group members in comparison to the unmitigated models. The domain-dependent impacts of model outcomes should help drive future bias mitigation method development.

KW - Fairness

KW - Impact

KW - Machine Learning

KW - Synthetic data

UR - http://www.scopus.com/inward/record.url?scp=85173617924&partnerID=8YFLogxK

U2 - 10.1145/3600211.3604699

DO - 10.1145/3600211.3604699

M3 - Conference paper

T3 - AIES 2023 - Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society

SP - 297

EP - 311

BT - AIES 2023 - Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society

PB - ACM

CY - Montreal, Canada

ER -