We present a new kind of support vector machine (SVM) by improving techniques arising from hyper sphere support vector machine (hs-SVM) for multi-class classification. First, we demonstrate the inaccuracy of original hs-SVM when serval hyper spheres have intersection. Next, we propose the relative radial distance as the classification basis, which can give better characterization on marginalization degree of the samples. Experiments quantitatively show the superior effectiveness as compared to the original hs-SVM on hand post...
