An Empirical Analysis on Data Preprocessing Over Two-Class versus Multi-Class Imbalance Learning

Abstract

Traditional artificial learning methodologies consider that the number of samples in each class is approximately same in size. But coming to real-time situations, instances distribution is uneven because some of class samples appear more frequent compare to others. This causes difficulty to learning algorithms which give favor to the majority class which has large no. of samples. This paper addresses useful methods related to data preprocessing on class imbalance problems and further this paper presents empirical analysis on data preprocessing techniques on binary class as well as multi class classification problems using evaluation metrics like Accuracy, AUC, G-Mean.