Global-local information based oversampling for multi-class imbalanced data

Multi-class imbalanced classification is a challenging problem in the field of machine learning. Many methods have been proposed to deal with it, and oversampling is one of the most popular techniques which alleviates class imbalance by generating instances for the minority classes. However, each oversampling utilizes a single way to generate instances for all candidate minority ones, which neglects the intrinsic characteristics among different minority class instances, and makes the synthetic instances redundant or ineffective. In this work, we propose a global-local-based oversampling method, termed GLOS. We introduce a new discreteness-based metric (DID) and distinguish the minority class from the majority class by comparing it with each class-level discreteness value. Then, for each minority class, some difficult-to-learn instances are selected, which have smaller instance-level dispersion than the corresponding class-level one, to generate synthetic instances. And the number of synthetic instances equals the difference between two types of dispersion values. These selected instances are assigned into different groups according to their local distribution. Furthermore, GLOS adopts a specific synthetic strategy to each group instance purposefully. Finally, all minority classes, part of the majority classes instances, and synthetic data will be used as training data. In this way, the quantity and quality of synthetic instances are guaranteed. Experimental results on KEEL and UCI data sets demonstrate the effectiveness of our proposal.