Boosting methods for multi-class imbalanced data classification: an experimental review

Since canonical machine learning algorithms assume that the dataset has equal number of samples in each class, binary classification became a very challenging task to discriminate the minority class samples efficiently in imbalanced datasets. For this reason, researchers have been paid attention and have proposed many methods to deal with this problem, which can be broadly categorized into data level and algorithm level. Besides, multi-class imbalanced learning is much harder than binary one and is still an open problem. Boosting algorithms are a class of ensemble learning methods in machine learning that improves the performance of separate base learners by combining them into a composite whole. This paper's aim is to review the most significant published boosting techniques on multi-class imbalanced datasets. A thorough empirical comparison is conducted to analyze the performance of binary and multi-class boosting algorithms on various multi-class imbalanced datasets. In addition, based on the obtained results for performance evaluation metrics and a recently proposed criteria for comparing metrics, the selected metrics are compared to determine a suitable performance metric for multi-class imbalanced datasets. The experimental studies show that the CatBoost and LogitBoost algorithms are superior to other boosting algorithms on multi-class imbalanced conventional and big datasets, respectively. Furthermore, the MMCC is a better evaluation metric than the MAUC and G-mean in multi-class imbalanced data domains.