An under-sampling imbalanced learning of data gravitation based classification

Abstract

With one class outnumbering another, many real classification tasks show imbalanced class distributions, which brings big trouble to standard classification models: they usually intend to recognize a minority instance as a majority one. The data gravitation based classification (DGC) model, a newly developed physical-inspired supervised learning model, has been proven effective for standard supervised learning tasks. However, DGC is not able to get high performances for imbalanced data sets, like most other standard learning algorithms do. Thus, to address the problem, an under-sampling technique, together with an ensemble technique, has been designed to adapt the standard DGC model for imbalanced learning tasks. The new adapted DGC model is called UI-DGC. 22 low imbalanced and 22 high imbalanced data sets are selected for the experimental study. UI-DGC is compared with standard and imbalanced learning algorithms. Empirical studies suggest that the UI-DGC model can get high imbalanced classification performances, especially for high imbalanced tasks.