An improved grid search algorithm of SVR parameters optimization

The proper selection of parameters, kernel parameter g, penalty factor c, non-sensitive coefficient p of Support Vector Regression (SVR) model can optimize SVR's performance. The most commonly used approach is grid search. However, when the data set is large, a terribly long time will be introduced. Thus, we propose an improved grid algorithm to reduce searching time by reduce the number of doing cross-validation test. Firstly, the penalty factor c could be calculated by an empirical formula. Then the best kernel parameter g could be found by general grid search algorithm with the achieved c and a p-value selected randomly within a range. According to the achieved c and p, the grid search algorithm is used again to search the best non-sensitive coefficient p. Experiments on 5 benchmark datasets illustrate that the improved algorithm can reduce training time markedly in a good prediction accuracy.