Análisis comparativo del rendimiento de algoritmos de clasificación binaria en un conjunto de datos desbalanceados

Abstract

Identifying the class of an object is a supervised machine learning task whose performance depends, almost exclusively, on the dataset used in training. Therefore, one of the challenges faced by classification algorithms, specifically binary classification, is learning to clearly distinguish between two classes, when you have a much larger number of instances of one class than another. To avoid bias in the classification, data balancing techniques are usually used that seek to balance the dataset, increasing or reducing the number of instances of the minority class and the majority class, respectively. This paper proposes a comparative analysis of the performance of four classifiers such as Logistic Regression, Random Forest, Artificial Neural Networks and Nayve Bayes combined with four different data balancing techniques Near Miss, SMOTE, SMOTEENN and SMOTETomek. The results show that Near Miss achieves a proper balance between the classes, so that the algorithms increased their overall performance, reaching precision and accuracy greater than 95\%. The rest of the techniques, on the other hand, did not increase the ability of the classifiers to identify objects of the minority class, with the exception of Random forest and Artificial Neural Networks, which achieved a true negative rate greater than 70%, while maintaining a true positive rate greater than 80%. Similarly, the training and testing times of the balanced data sets with oversampling techniques or hybrids are far superior to the times obtained by undersampling techniques such as Near Miss, since the latter reduces the number of instances to be processed by the models.