Detecção de fraudes em cartões de crédito utilizando métodos de baseados em árvores de decisão

Abstract

In 2020, the world was surprised by the arise of the pandemic. Due the social isolation issued by this, Internet purchases had a significant increase, and, with them, fraud attempts has increased as well, especially in credit card purchases. Although the fraud increased, the legit transactions are still expressively bigger, which difficult its detection. With this in mind, is known that fraudulent transactions are considered outliers towards the legit ones. This work aims to utilize tree-based machine learning algorithms, which are easily interpretable models for humans and can be used to detect potential frauds. The dataset was split into training, testing, and validation sets in a stratified manner. The testing set was generated using the largest fraction of the data to create representative models, meaning that a smaller portion of the data was used for training. During the model construction, the data was standardized using the StandardScaler before proceeding to the next steps, which were hyperparameter optimization with BayesianSearch, training the Isolation Forest and ExtraTree models with the found hyperparameters, prediction, and K-fold cross-validation. Finally, to compare the models, the Friedman statistical hypothesis test was applied with a significance level of 95%. Since an extremely significant statistical difference was found, the Nemenyi post-test was applied to determine which pairs of models had a statistically significant difference. As a result, it can be concluded with 95% certainty that the ET models performed better compared to the IF-Matt model. Additionally, the supervised model achieved better classification measures for legitimate transactions, while the unsupervised model excelled in classifying frauds. It can also be observed that models with higher sensitivity had a higher Matthews correlation coefficient.