Samples Selection Based on SVR for Prediction of Steel Mechanical Property

Support Vector Regression is a new kind of machine learning algorithm based on the idea of structural risk minimization with good generalization performance, which is applied to build prediction model for steel mechanical property in this paper. Training SVR requires large memory and long CPU time when the data set is large. To alleviate the computational burden in SVR training, a new sample selection algorithm is proposed, which calculate the times for bootstrap samples locating outside the tube and decide those samples with larger probability according to the times as selected samples for modeling. Simulation result and the performance of practical application in some steel factory show that the proposed algorithm reserve effective samples, and also improve the performance of the SVR modeling.