Abstract

A weighting strategy for handling outliers in binary classification using support vector machine (SVM) is proposed in this article. The traditional SVM model is modified by introducing an induced ordered weighted averaging (IOWA) operator, in which the hinge loss function becomes an ordered weighted sum of the SVM slack variables. These weights are defined using IOWA quantifiers, while the order is induced via fuzzy density-based methods for outlier detection. The proposal is developed for both linear and kernel-based classification using the duality theory and the kernel trick. Our experimental results on well known benchmark datasets demonstrate the virtues of the proposed IOWA-SVM, which achieved the best average performance compared to other machine learning approaches of similar complexity.