Online Lagrangian Support Vector Regression

To quickly solve online support vector regression algorithm, an algorithm of online incremental learning which based on Lagrangian support vector regression (LSVR) was proposed. Unconstrained optimization problems of LSVR could be solved by a rapid iterative algorithm. It could converge from any starting point. LSVR had the advantage that its solution was obtained by taking the inverse of a matrix of order equaled to the number of input samples plus one at the beginning of the iteration. The incremental learning algorithm of LSVR was presented by using the S-M-W identity to reduce the computing time in the linear case. As for the nonlinear case, the inversion of the matrix after increment was solved based on the previous computed information and it was unnecessary to repeat the computing process. The effectiveness of the proposed method was illustrated with several data sets. Results show that the proposed algorithm compared with the previous algorithm not only maintains better accuracy, but also greatly shortens the training time.