Approximately Linear Dependency between μ and the Input Noise in Huber-Support Vector Regression
Xiao Tao Zhou · 2007
The dependency relationship between μ and the input noise in Huber-SVR is studied using SVR Bayesian evidence framework. First, focus is paid on the cases of Laplacian noise and Uniform noise, and the approximately linear dependencies between μ and the variances of the two noises are then respectively derived. Second, with the relevant conclusion on Huber-SVR and experimental study, the more general claim is then proposed that the approximately linear dependency is almost kept between μ and the input noise in Huber-SVR. Such a dependency relationship is useful to determine the optimal choice for μ in Huber-loss function in the existence of unknown input noise.