Catálogo do requisito não-funcional justiça para aplicações de aprendizado de máquina

Abstract

Machine learning (ML) is a branch of artificial intelligence, where a computer has the ability to learn from experience, using data for its learning. It is an area that has attracted a lot of attention, however ensuring safe, fair, transparent and accurate ML-enabled systems is a challenge. These are some important quality characteristics for ML. Quality characteristics are a type of non-functional requirement (NFR), that are important for the expected behavior and quality of systems. Fairness is one of these NFRs, because when Fairness is not considered a priority, it results in systems with unexpected and biased behavior. Considering the Softgoal Interdependence Graph (SIG), which is a known format for cataloging NFRs, this work aims to capture sub-characteristics for Fairness and catalog them in a SIG. An existing approach was used to catalog the NFR, called ARRANGE, using a well-defined research method: grounded theory. As a final result, a Justice SIG was obtained, consisting of two main subcategories and nine subcharacteristics. Therefore, software engineers, whenever they need to, can and should consult this catalog to help them make decisions regarding the use and application of this NFR, for ML systems that consider Fairness as an NFR.