Análise de transformações metamórficas em conjunto de dados para garantia de fairness em modelos de classificação de informações.

Abstract

Classification bias is a recurring problem in learning systems and is also caused by the presence of real-world prejudices and injustices embedded in digital data. The study of this subject focuses on the area known as Fairness, which, although it does not have a unified definition in the literature, represents the assurance that decisions made by systems are carried out impartially, avoiding the presence of prejudice and discrimination against minorities of social groups, whether by gender, race, purchasing power, etc. However, such systems are often labeled as ‘untestable’, since it is not clear what the correct results for data classifications are, and this raises the discussion about how models can generalize their results. Aiming for impartiality from the generalization of the models, Metamorphic Transformations techniques emerge, with changes made to the datasets, both in their structures and also in their values. These techniques are used in this work with the aim that the classifications of the models executed with the altered data can generate the most impartial results possible, regardless of the social groups present in the data. The Metamorphic Transformations employed are conceptualized and executed in four different scenarios, either by changing information in parts of the values of each attribute of the dataset used or by changing all information from a single group of samples, such as information about people of just one sex, male or female. It is within this context that this work is developed, through the analysis of different applications and combinations of metamorphic transformations in sample data that have the female gender as a group susceptible to prejudice on the part of classification models. Their values are highlighted through the calculation of metrics focused on the areas of Fairness, carrying out the analysis not only with the classified values but also taking into account the real values, in addition to the calculation of consolidated metrics vi vii vii in Information Classification tasks. In terms of results, it was possible to achieve values that indicate improvements of up to 20% when using model instances trained with the transformed data. By analyzing the transformations in different learning models, it was possible to support the discussion on whether it is really possible to improve Fairness indices in classifications, with some scenarios showing promise in answering this question and also in expanding the literature on the subject, which until then has proven to be scarce.