Interpretabilidade de modelos de machine learning: aplicação no mercado de crédito

Abstract

In the credit market, it is common to use the information available to establish default risks for a potential customer, and to obtain good statistical accuracy in predicting this risk, machine learning techniques are employed in order to use the data in producing a score, known as a credit score. In this work, four techniques are evaluated - Logistic Regression, Decision Tree, Gradient Boosting Machine and Multilayer Perceptron - in the concept of credit granting, from the perspective of interpretability of the construction of the score and performance. The objective is to present the way the models use the data to make decisions and show this information in a coherent way to third parties, even when a model considered "black box" is applied. Finally, texts are generated that justify the composition of the score attributed to a specific customer, in order to observe its consistency with business usage.