A comparative analysis between undersampling and oversampling approaches to data balancing
Ronaldo Agra, Fernando Narciso Bertolaccini De Souza, Paulo Célio Soares, Gladston Luiz da Silva · 2023
The constant increase in fraudulent operations in credit card transactions has led financial institutions to prioritize the use of machine learning models in fraud detection. Challenging task due to the large volume of data and the disproportion between true and fraudulent transactions. In this sense, data imbalance is a critical factor in the construction of models and, therefore, it is important to balance the classes of “fraud“ and “non-fraud“ in the pre-processing of the data. This work aimed to carry out a comparative analysis between two data balancing approaches, from a public database that represents credit card transactions. An experiment was performed to compare simple random subsampling without replacement and oversampling by SMOTE in the data balancing task. The datasets were used in the construction of two logistic regression models to classify transactions into “fraud“ and “non-fraud”. It was concluded that both models are equivalent in accuracy, but the one that used the SMOTE oversampling technique showed lower percentages of false positives and negatives.