Modelo de classificação para dados desbalanceados: método SMOTE e variantes

Abstract

Often, in classification models, we encounter databases that have highly imbalanced classes, such as: rare disease diagnostic data, manufacturing defects, fraudulent transactions, etc. Training a model on a dataset with few observations of a particular class results in poor predictive performance, especially for observations belonging to the minority class. In this Undergraduate Thesis, we present and compare different variants of the Synthetic Minority Over-sampling TEchnique (SMOTE) method for oversampling imbalanced data used in classification models, specifically Logistic Regression, in order to demonstrate how these techniques can improve the ability to identify and predict observations from the minority class in realistic and imbalanced scenarios, as well as to determine which combination of sampling technique and Logistic Regression classification model leads to better performance.