Global Geometry of SVM Classifiers
Dengjian Zhou, Zhou Xiao B, Rujia Dai · 2002
We construct an alternative geometry framework for Support Vector Machine (SVM) classifiers. Within this framework, separating hyperplanes, dual descriptions and solutions of SVM classifiers all are constructed clearly by a pure geometry fashion. Now all kinds of SVM formulations and their dual descriptions including the arbitrary-norm cases are only different expressions of the underlying common geometry essentials. Compared with the optimization theory in SVM classifiers, we don't need redundant confused computations any more. Instead, every step in our theory is guided by elegant geometry intuitions. Our framework can make people understand SVM in a totally visual fashion. In addition, it is also helpful to expose the correlations between SVM and other learning algorithms.