Um estudo comparativo de métodos para balanceamento do conjunto de treinamento em aprendizado de redes neurais artificiais

Abstract

Studies in supervised learning have shown that classifiers induced from imbalanced data sets have presented a reduced performance. A intuitive solution for this problem is to balance the training set. This work presents a study of resampling methods for balancing training sets of neural networks. Particularly, five methods were selected and tested: Smote, Smote + ENN, Smote + Tomek Link, BED and ADASYN. The results obtained were analyzed using metrics from ROC analysis.