Using Information on Class Interrelations to Improve Classification of Multiclass Imbalanced Data: A New Resampling Algorithm

The relations between multiple imbalanced classes can be handled with a specialized approach which evaluates types of examples’ difficulty based on an analysis of the class distribution in the examples’ neighborhood, additionally exploiting information about the similarity of neighboring classes. In this paper, we demonstrate that such an approach can be implemented as a data preprocessing technique and that it can improve the performance of various classifiers on multiclass imbalanced datasets. It has led us to the introduction of a new resampling algorithm, called Similarity Oversampling and Undersampling Preprocessing (SOUP), which resamples examples according to their difficulty. Its experimental evaluation on real and artificial datasets has shown that it is competitive with the most popular decomposition ensembles and better than specialized preprocessing techniques for multi-imbalanced problems.