Proximal Support Vector Machine Classifier based on LMS Algorithm

In proximal support vector machine (PSVM), datapoints are assigned to classes by measuring the proximity of it from the two parallel hyperplanes. The hyperplanes are formed such that each of the two parallel hyperplanes is closest to only one of the datasets out of two datasets and the two parallel hyperplanes should be as far as possible from each other. The benefit of PSVM is that it demands less training time as the size of the training patterns increases and less computational complexity as compared to standard SVM. In this paper, an improved version of proximal support vector machine (PSVM) is proposed for pattern classification of binary datasets. In the proposed PSVM technique, the weight vector is modified based on least mean square (LMS) algorithm, which reduces the training classification error. The idea behind the proposed technique is to enlarge the separating boundary, such that the data points of each class lie on the correct side of the hyperplane and the separability between the datasets is increased. The performance of the proposed method is evaluated on several benchmark datasets for both linear and nonlinear classifiers. Experimental results show that the proposed LMS based proximal support vector machine (LMS-PSVM) classifier performs better compared to standard SVM and PSVM.