TY - GEN

T1 - Multi-fairness Under Class-Imbalance

AU - Roy, Arjun

AU - Iosifidis, Vasileios

AU - Ntoutsi, Eirini

N1 - Funding Information: Acknowledgements. The work of the first author is supported by the Volkswagen Foundation under the call “Artificial Intelligence and the Society of the Future” (the BIAS project). We are sincerely thankful to the invaluable suggestion of Prof. Niloy Ganguly from L3S Research Center, in shaping up the paper to its current form. Most of the work was carried out while the last author was affiliated with Freie Universität Berlin, Germany.

PY - 2022

Y1 - 2022

N2 - Recent studies showed that datasets used in fairness-aware machine learning for multiple protected attributes (referred to as multi-discrimination hereafter) are often imbalanced. The class-imbalance problem is more severe for the protected group in the critical minority class (e.g., female +, non-white +, etc.). Still, existing methods focus only on the overall error-discrimination trade-off, ignoring the imbalance problem, and thus they amplify the prevalent bias in the minority classes. To solve the combined problem of multi-discrimination and class-imbalance we introduce a new fairness measure, Multi-Max Mistreatment (MMM), which considers both (multi-attribute) protected group and class membership of instances to measure discrimination. To solve the combined problem, we propose Multi-Fair Boosting Post Pareto (MFBPP) a boosting approach that incorporates MMM-costs in the distribution update and post-training, selects the optimal trade-off among accurate, class-balanced, and fair solutions. The experimental results show the superiority of our approach against state-of-the-art methods in producing the best balanced performance across groups and classes and the best accuracy for the protected groups in the minority class.

AB - Recent studies showed that datasets used in fairness-aware machine learning for multiple protected attributes (referred to as multi-discrimination hereafter) are often imbalanced. The class-imbalance problem is more severe for the protected group in the critical minority class (e.g., female +, non-white +, etc.). Still, existing methods focus only on the overall error-discrimination trade-off, ignoring the imbalance problem, and thus they amplify the prevalent bias in the minority classes. To solve the combined problem of multi-discrimination and class-imbalance we introduce a new fairness measure, Multi-Max Mistreatment (MMM), which considers both (multi-attribute) protected group and class membership of instances to measure discrimination. To solve the combined problem, we propose Multi-Fair Boosting Post Pareto (MFBPP) a boosting approach that incorporates MMM-costs in the distribution update and post-training, selects the optimal trade-off among accurate, class-balanced, and fair solutions. The experimental results show the superiority of our approach against state-of-the-art methods in producing the best balanced performance across groups and classes and the best accuracy for the protected groups in the minority class.

KW - Boosting

KW - Class-imbalance

KW - Multi-discrimination

UR - http://www.scopus.com/inward/record.url?scp=85142727007&partnerID=8YFLogxK

U2 - 10.48550/arXiv.2104.13312

DO - 10.48550/arXiv.2104.13312

M3 - Conference contribution

AN - SCOPUS:85142727007

SN - 9783031188398

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 286

EP - 301

BT - Discovery Science

A2 - Pascal, Poncelet

A2 - Ienco, Dino

T2 - 25th International Conference on Discovery Science, DS 2022

Y2 - 10 October 2022 through 12 October 2022

ER -