Evolutionary inversion of class distribution in overlapping areas for multi-class imbalanced learning

Abstract

Inductive learning from multi-class and imbalanced datasets is one of the main challenges for machine learning. Most machine learning algorithms have their predictive performance negatively affected by imbalanced data. Although several techniques have been proposed to deal with this difficulty, they are usually restricted to binary classification datasets. Thus, one of the research challenges in this area is how to deal with imbalanced multiclass classification datasets. This challenge become more difficult when classes containing fewer instances are located in overlapping regions of the data attribute space. In fact, several studies have indicated that the degree of class overlapping has a higher effect on predictive performance than the global class imbalance ratio. This paper proposes a novel evolutionary ensemble-based method for multi-class imbalanced learning called the evolutionary inversion of class distribution in overlapping areas for multi-class imbalanced learning (EVINCI). EVINCI uses a multiobjective evolutionary algorithm (MOEA) to evolve a set of samples taken from an imbalanced dataset. It selectively reduces the concentration of less representative instances of the majority classes in the overlapping areas while selecting samples that produce more accurate models. In experiments performed to evaluate its predictive accuracy, EVINCI was superior to state-of-the-art ensemble-based methods for imbalanced learning.