O impacto da seleção de atributos e da separação dos dados de detecção de fraudes na tarefa de classificação

Abstract

In the context of credit card transaction fraud, it is necessary to enhance anti-fraud systems by considering the data splitting strategy and feature selection. This study evaluates how feature selection and data separation affect the performance of machine learning models. Three feature selection strategies and five training approaches were applied using three tree-based models. Techniques such as K-Means, undersampling, and nested cross-validation were combined to create different experimental scenarios. Feature selection II achieved good and consistent results across all experiments.