Maximal margin multi-classifier based on SVM hyperparameter tuning

Multiclass classification is the task of classifying the samples into more than two classes. Generally multi-classifiers face difficulty in classifying samples those are very close to the separating hyperplane, known as Generalization error. Generalization error can be reduced by maximizing the margin of the separating hyperplanes. Support Vector Machine (SVM) is a maximum-margin classifier, its aim is to maximize the width of the margin between the classes. An important step in the construction of SVM is to select optimal hyperparameters (kernel parameter). The kernel parameter in the kernel function affects the performance of the classifier, so the kernel parameter should be properly tuned. The quasi-Newton method is used to tune the kernel parameter, which maximizes the distance between classes. Experimental results on the real world data (linear, nonlinear and multiclass) shows that the proposed method gives better accuracy by maximizing the distance between the classes.