An improved model of support vector machine for high noise data regression problem

Abstract

For regression problem for high noise data by support vector machine, if the ɛ in ɛ-insensitive loss function takes the large value, the error will be big; if the ɛ takes the small value, the number of support vectors will be big, and the applications of support the vector machine will be influenced. To solve this problem, a method is brought to simplify the support vector machine based on the minimize maximal-error idea. Experiments show that this new support vector machine not only reduces the number of support vectors and regression time but also the number of support vectors do not obviously increase when the ɛ reduces. Therefore a theoretical foundation is established to structure the support vector machine for regression with high accuracy and low complicacy degree.