Support vector machine regression algorithm based on chunking incremental learning

Abstract

On the basis of least squares support vector machine regression (LSSVR), an adaptive and iterative support vector machine regression algorithm based on chunking incremental learning (CISVR) is presented in this paper. CISVR is an iterative algorithm and the samples are added to the working set in batches. The inverse of the matrix of coefficients from previous iteration is used to calculate the regression parameters. Therefore, the proposed approach permits to avoid the calculation of the inverse of a large-scale matrix and improves the learning speed of the algorithm. Support vectors are selected adaptively in the iteration to maintain the sparseness. Experimental results show that the learning speed of CISVR is improved greatly compared with LSSVR for the similar training accuracy. At the same time the number of the support vectors obtained by the presented algorithm is less than that obtained by LSSVR greatly. © Springer-Verlag Berlin Heidelberg 2006.