Justiça em aprendizagem de máquina na estimativa de risco de contratos públicos.

Abstract

Brazilian government signs contracts with companies for products acquirement and services provision. However, due to high demand for contracts audition, control units need to prioritize these contracts, which are usually done through risk estimation of contracts or companies. With its success in other contexts, machine learning techniques have been used in risk estimation by this agencies. Meanwhile, recent works have shown that decision making systems similar to machine learning risk estimation can be unfair. This observation points out the risk that machine learning models created by control units may be unfair. This Master’s thesis presents an assessment of justice in risk estimation models similar to those used by Brazilian federal and state units, using databases available for these units. The risk estimation studied models include both ad-hoc methods available in agencies and machine learning method, using a risk estimation methodology analogous to one published in an article by a federal control unit. Furthermore, three state-of-the-art methods were applied with the objective of mitigate injustices by the risk estimation models. Our results show that young companies are more falsely accused compared with the consolidated ones in all scenarios, besides others sensible classes in specific contexts. Furthermore, there wasn’t a consensus in which mitigation method had the best performance. Despite that, in all studied scenarios there was a model that improves justice at least in one sensible class. Half of the scenarios had an improvement of efficiency and justice, but the other half had a trade-off between justice and efficiency. In this way, it is expected that control units can pay attention to injustice present in risk estimation model and provide a method to mitigate them.