An improved ensemble learning method based on SVR

As a universal learning method, Support Vector Regression (SVR) has strong generalization ability and can perfectly solve some practical problems, such as small samples, nonlinear, high dimension and so on. However, the prediction accuracy of a single SVR is limited. With the help of ensemble learning, the sample prediction accuracy of SVR can be effectively improved. In ensemble learning, the construction method of training samples is a key. The larger difference between the training sample sub-sets leads to the stronger generalization ability of SVR. In this work, an improved method of sample construction is proposed to increase the differences between training sample sets. The proposed method divides the samples into several sub-categories by clustering algorithm. Each sub-category adds the samples closing to the clustering center from other sub-categories to form a new training sample set. The experiment results demonstrate that the proposed improved method has less number of iterations and higher predict accuracy as compared to the method with random sampling.