A New SVM Considering the Sample Distribution

Abstract

As the continuation of former work, a new SVM is discussed. It is similar to the Fisher Discriminant Analysis (FDA) that the normal vector of SVM separating hyperplane is a one-dimension vector in the high-dimension hidden space, which can be used as a projection orientation to classify data. So, after the SVM training, the projection of samples can be calculated by kernel function. Finally, the threshold of the classifier is ascertained according to the distribution of projection. With the sample distribution considered, SVM performance is improved. Simulation manifests the validity of the method in this paper.