Dynamic Centroid Insertion and Adjustment for Data Sets with Multiple Imbalanced Classes

The imbalance problem is receiving an increasing attention in the literature. Studies on binary cases are recurrent but limited when considering the multiple classes approach. Solutions to imbalance domains may be divided into two groups, data level approaches, and algorithmic approaches. The first approach is more common and focuses on changing the training data aiming to balance the data set, oversampling the smallest classes, undersampling the biggest ones or using a combination of both. Instance reduction is another approach to the problem. It tries to find the best-reduced set of instances that represent the original training set. In this work, we propose a new Prototype Generation method called DCIA. It dynamically inserts new prototypes for each class and then adjusts their positions with a search algorithm. The set of generated prototypes may be used to train any classifier. Experiments showed its potentiality by enabling an INN classifier to perform sometimes as well or even better than some ensemble classifiers created for different multiclass imbalanced domains.