Abordagem para verificação autoexplicativa de erros em classificadores

Abstract

The relevant role, in different areas of society, of Artificial Intelligence and Machine Learning systems is perceptible, especially when used in conjunction with Decision Support Systems. However, the lack of transparency and explainability of these systems has been a frequent concern. Explainable Artificial Intelligence emerges in this scenario as a solution to this problem, enabling intelligent systems to gain the trust of their users. The approach of this study is relevant to improve such systems in critical areas, boosting more reliable and transparent decision-making. Therefore, this work proposes an approach to verify and explain errors made by classifiers, contributing to confidence and wider adoption of artificial intelligence applications. In the bases adopted in this study, the results obtained demonstrated the efficiency of the proposal in correcting all errors in one base, 75% in another and 66% in two others, leaving only three bases uncorrected out of a total of nine bases.