Support Vector Machine Classification using Mahalanobis Distance Function

Support Vector Machine (SVM) is a powerful technique for data classification. The SVM constructs an optimal separating hyper-plane as a decision surface, to divide the data points of different categories in the vector space. The Kernel functions are used to extend the concept of the optimal separating hyper-plane for the non-linearly separable cases so that the data can be linearly separable. The different kernel functions have different characteristics and hence the performance of SVM is highly influenced by the selection of kernel functions. Thus, despite its good theoretical foundation, one of the critical problems of the SVM is the selection of the appropriate kernel function in order to guarantee high accuracy of the classifier. This paper presents the classification framework, that uses SVM in the training phase and Mahalanobolis distance in the testing phase, in order to design a classifier which has low impact of kernel function on the classification accuracy. The Mahalanobis distance is used to replace the optimal separating hyperplane as the classification decision making function in SVM. The proposed approach is referred to as Euclidean Distance towards the Center (EDC_SVM). This is because the Mahalanobis distance from a point to the mean of the group is also called as Euclidean distance towards the center of data set. We have tested the performance of EDC_SVM on several datasets. The experimental results show that the accuracy of the EDC_SVM classifier to have a low impact on the implementation of kernel functions. The proposed approach also achieved the drastic reduction in the classification time, as the classification of a new data point depends only on the mean of Support Vectors (SVs) of each category.