G-SOMO: An oversampling approach based on self-organized maps and geometric SMOTE

Traditional supervised machine learning classifiers are challenged to learn highly skewed data distributions as they designed expect classes equally contribute the minimization of cost function. Moreover, design expects equal misclassification costs, causing a bias for overrepresented classes. Different strategies have been proposed correct this issue. The modification set has become common practice since procedure is generalizable all classifiers. Various algorithms rebalance distribution through creation synthetic instances were in past. In paper, we propose new oversampling algorithm named G-SOMO. identifies optimal areas create artificial an informed manner and utilizes geometric region during generation process increase their variability. Our empirical results on 69 datasets, validated with different metrics against benchmark commonly used methods show that G-SOMO consistently outperforms competing methods. Additionally, statistical significance our established.