Abstract:

The influence of the local support vector on the prediction results is not fully considered in the traditional ε insensitive support vector regression (ε-SVR), which is not conducive to improve the predictive accuracy of regression problems. An error correction method is proposed for ε-SVR, in which the minimum sum of Euclidean distances between ideal values and ε-SVR regression values and local support vectors are taken as the objective function, and the width of εinsensitive loss tube (εtube) is taken as constraint to correct the error in terms of local support vector on and out of the ε tube boundary in high dimensional feature space. Simulation using artificial datasets with different distributed and UCI benchmark data sets shows that the proposed method has higher prediction and generalization performance.