GWSVM: A Robust GMM-Based Pattern Weighting for SVM

Support Vector Machine (SVM) is an efficient classifier due to its generalization capability; however, it suffers from high sensitivity to noisy data. A new weighted SVM classifier is proposed here to overcome this problem by a pattern weighting approach based on Gaussian Mixture Models (GMM), called hereinafter GWSVM. The weight learning is a procedure of knowledge extraction from the feature space. First, a clustering phase is carried out according to Gap Statistic and Subtractive Clustering (SC) methods. Then a GMM is trained using the clustered data and pattern weights are determined based on the GMM outputs. In the third phase, a redefined optimization formula regarding the weighted patterns is used, so SVM decision boundary is adjusted sensibly. GWSVM is applied to some UCI repository datasets and its performance is presented. The results show the supremacy of our novel classifier in comparison with other versions of SVM. Further, to evaluate robustness and generalization power of GWSVM, the impact of white noise with different variances is investigated. Applying GWSVM and SVM classifiers to noisy data manifests effectiveness of GWSVM in noisy environment.