Justiça em modelos de inteligência artificial: um estudo comparativo de técnicas de mitigação e suas implementações no contexto educacional

Abstract

Machine learning models are increasingly being used to influence decision-making that directly and indirectly impacts people's lives. Specifically, decision-making related to education can have a profound influence, both in the present and in the future, on the repercussions of the results generated by these models. In this context, there has been growing concern about the fairness of models, that is, whether they have biases that are incompatible with the equity of socially stigmatized groups, such as gender, race, age, income, etc. The objective of this work is to compare the main libraries that implement fairness metrics and mitigation techniques, and to provide data that help choose the most appropriate mitigation implementation for similar cases, in addition to demonstrating the impact of sensitive variables on the output of the models.