Comparação de Métodos de Balanceamento de Dados em Diferentes Cenários

Abstract

Class imbalance occurs when there is an unequal distribution among the classes in a dataset, such that one class has a significantly smaller number of instances compared to the others. This scenario represents one of the challenges present in machine learning tasks, directly impacting the performance of predictive models. This study aims to compare different data balancing techniques, including oversampling, undersampling, and hybrid approaches, applied to multiple datasets with different levels of imbalance. Experiments were conducted using classification algorithms such as Decision Tree (C5.0), Random Forest, Artificial Neural Networks (ANN), Logistic Regression, and SVM, which were evaluated using the following performance metrics: accuracy, precision, recall, and F1-score. The analysis of the results made it possible to identify the impact of balancing techniques in different experimental scenarios, as well as their limitations and advantages. In general, oversampling techniques frequently showed superior or equivalent performance compared to the other evaluated methods, while the use of unbalanced data tended to present inferior results for certain metrics and algorithms. In addition, tree-based models, such as Random Forest, demonstrated greater robustness across the different analyzed scenarios. The findings of this study aim to provide support for the appropriate selection of balancing methods, contributing to the development of more robust and reliable models.