The Multi-attribute Fairer Cover Problem

• Ana Paula S. Dantas UNICAMP https://orcid.org/0000-0002-8831-0710
• Gabriel Bianchin de Oliveira UNICAMP https://orcid.org/0000-0002-1238-4860
• Helio Pedrini UNICAMP https://orcid.org/0000-0003-0125-630X
• Cid C. de Souza UNICAMP https://orcid.org/0000-0002-5945-0845
• Zanoni Dias UNICAMP https://orcid.org/0000-0003-3333-6822

Resumo

Alongside the increased use of algorithms as decision making tools, there have been an increase of cases where minority classes have been harmed. This gives rise to study of algorithmic fairness that deals with how to include fairness aspects in the design of algorithms. With this in mind, we define a new problem of fair coverage called Multi-Attribute Fairer Cover, that deals with the task of selecting a subset for training that is as fair as possible. We applied our method to the age regression model using instances from the UTKFace dataset. We also present computational experiments for an Integer Linear Programming model and for the age regression model. The experiments showed significant reduction on the error of the regression model when compared to a random selection.