Techniques to deal with imbalanced data in multi-class problems: A review of existing methods
Vitor Miguel Saraiva Esteves · Open Repository of the University of Porto (University of Porto) · 2020
Imbalanced learning is one of the most challenging classification problems in the field of machine learning and has been receiving increased attention over the years due to the necessity of handling real world information which is usually skewed.This effect occurs when one of the classes has a bigger number of examples compared to the rest.When we attempt to classify data in said conditions, machine learning algorithms will be able to correctly identify the majority class examples but will most likely fail when attempting to identify minority class examples, which often end up representing the most valuable information.Several surveys were published in the last few years about approaches to solve the problem.Adding to this issue, when we have a multi-class scenario where the examples to be classified can fall into more than two classes, the accuracy of applied techniques plummets due to their inability to deal with the issue.In the latest years, various techniques were proposed to deal with the matter, and they usually do so by converting the problem into subsets of two-class problems that can be solved by common classifiers.This process is called class decomposition.However, recent studies show that it can cause class overlapping and loss of valuable information.In this study, we address the topic first, by identifying algorithms that deal with multi-class imbalance without using class decomposition and categorizing them based on their approach.Then, we proceed to benchmark the two latest state-of-the-art ensemble algorithms, MBBR and SOUPBagging against each other and see how they perform in 9 real life datasets.The results showcase their ability to handle multi-class imbalance with high accuracy for classifying skewed data.Lastly, a possible future direction in the field is briefly discussed.