Aplicação de um modelo de aprendizado de máquina enviesado: detecção de fraudes em transações de cartão de crédito no comércio eletrônico

Abstract

The case study occurs in a context of high volume of fraud attempts in e-commerce, specifically with credit cards. The work seeks to create a machine learning model that aims to hit the positive class, acting as an intermediary to avoid the use of models offered in the fraud prevention market. The applied methodology consisted of data processing, selection of the best variables, correction of minority class imbalance, application of Random Forest, Logistic Regression and Gradient Boosting algorithms, tuning algorithms’ hyperparameters, selection of threshold and creation of ranges with suggestions of greater or lesser exposure to risk. As a result, it was possible to observe a low volume of error in the positive class and the possibility of avoiding the use of up to 80% of the market model studied.