Data stratification analysis on the propagation of discriminatory effects in binary classification
Diego Minatel, Angelo Cesar Mendes da Silva, Nícolas Roque dos Santos, Mariana Cúri, Ricardo Marcondes Marcacini, Alneu de Andrade Lopes · 2023
Unfair decision-making supported by machine learning, which harms or benefits a specific group of people, is frequent. In many cases, the models only reproduce the biases in the data, which does not absolve its responsibility for these decisions. Thus, with the increase in the automation of activities through machine learning models, it is mandatory to prospect solutions that add fairness factors to the models and clarity about the supported decisions. One option to mitigate model discrimination is quantifying the ratio of instances belonging to each target class to build data sets that approximate the actual data distribution. This alternative aims to reduce the responsibility of data on discriminatory effects and direct the function of treating them to the models. In this sense, we propose to analyze different types of data stratification, including stratification by sociodemographic groups that are historically unprivileged, and associate these stratification types to the fairer or unfairer models. According to our results, stratification by class and group of people helps to develop fairer models, reducing the discriminatory effects in binary classification.