Effective Approaches to Improve the Classification Performance in Multiclass Imbalanced Datasets

Abstract

Many real-world classification datasets exhibit class imbalance, characterized newlineby a disproportionate distribution of examples among different class labels in the newlineproblem. This disproportion poses a challenge to classifier performance as they are newlinetypically designed with a focus on accuracy, leading to the neglect of minority newlineclasses. Handling class imbalances is particularly challenging in multi-class newlinescenarios and becomes more pronounced when imbalance is present. In situations newlineinvolving multiple majority and multiple minority classes, determining ahead which newlineclasses should be emphasized during the learning phase is not straightforward, newlineunlike in binary problems. The techniques developed for imbalanced binary newlineclassification often do not directly apply to scenarios with multiple classes. newlineFurthermore, assessing the effectiveness of learning models in the presence of an newlineuneven distribution of samples among multiple classes poses a considerable newlinechallenge. With metrics such as accuracy, the overrepresentation of classes with a newlinegreater number of samples may conceal inadequate classification performance in the newlineless prevalent classes. Considerable attention and solutions have been directed newlinetoward resolving the binary or two-class imbalance issue. However, there is limited newlineattention given to multi class-imbalanced datasets that involve more than two class newlinelabels with changing degrees of imbalance. In addition to the diminished newlinerepresentation of instances, several factors can impact the predictive performance of newlinetraditional classifiers on minority instances. These factors include noise, overlapping newlineclasses, multi-dimensional classes, and instances that exhibit characteristics of both newlinemulti-minority and multi-majority classes. Therefore, there is motivation to devise newlineapproaches that effectively handle the recognition of the minority class as its primary newlineobjective.