Improved slack factor based fuzzy support vector machine

Abstract

Class imbalanced datasets broadly exist in real world problems; the support vector machine (SVM) algorithm is inecient in dealing with imbalance problems, noise, and outliers. Fuzzy support vector machines (FSVMs), where di↵erent fuzzy memberships are defined for di↵erent samples, have been successfully employed for class imbalance learning. However, fuzzy membership function is generally impaired by imbalanced datasets, inferring the incorrect measure of the sample’s importance and altering the performance of FSVMs. A new slack-factor-based fuzzy support vector machine (SFFSVM) model has been proposed to address this issue. Based on the slack factor, SFFSVM defines fuzzy membership considering the positional link between optimal and estimated hyperplanes. SFFSVM assigns large membership to those wrongly classified majority class samples having inappropriately high, i.e., 2 slack factor value leading to incorrect classification of correctly classified minority samples by the hyperplane obtained by Di↵erent Error Cost (DEC). To overcome this problem, we develop an improved slack-factor-based fuzzy support vector ma chine (ISFFSVM), introducing a new parameter named the location parameter. The benefit of our model is that the DEC hyperplane cannot go beyond the position where the slack factor score of the majority class observations approaches the value of the location parameter. Consequently, more minority-class samples are correctly classified in the proposed ISFFSVM model as compared to SFFSVM model. It is evident from the thorough tests on real-world KEEL datasets that the proposed ISFFSVM performs better than the baseline classifier.