SVM ensemble training for imbalanced data classification using multi-objective optimization techniques

Abstract One of the main problems with classifier training for imbalanced data is defining correct learning criterion. On one hand, we want minority class to be correctly recognized, and on other do not make too many mistakes in majority class. Commonly used metrics focus either predictive quality distinguished or propose an aggregation simple metrics. The aggregate metrics, such as Gmean AUC , are primarily ambiguous, i.e., they indicate specific values errors made Additionally, improper use results solutions selected their help that may favor authors realize a solution this problem using overall risk. However, requires knowledge costs associated between classes, which often unavailable. Hence, paper will semoos algorithm - approach based multi-objective optimization optimizes criteria related prediction both classes. returns pool non-dominated from user can choose model best suits him. Automatic selection formulas so-called Pareto front have also been proposed compare state-of-the-art methods. train svm ensemble dedicated classification task. experimental evaluations carried out large number benchmark datasets confirm its usefulness.