The Comparison of the Impact of Two ε- Insensitive Loss Functions in Support Vector Regression

Abstract

In Support Vector Regression(SVR), two loss functions have been proposed : the ε-insensitive loss function and the quadratic ε-insensitive loss function. However, the latter has only been proposed, and only linear SVR has been implemented in programs. In this study, we implemented a program for nonlinear SVR using the quadratic ε-insensitive loss function and compared its prediction accuracy with that of SVR using the ε-insensitive loss function. As a result, it was observed that the SVR using the quadratic ε-insensitive loss function has smaller maximum squared error and maximum absolute error between the predicted value and the observed value than the SVR using the ε-insensitive loss function.