|
In least squares support vector machine (LSSVM), nonlinear function estimation is done by solving a linear set of equations instead of solving a quadratic programming problem, and a nonsparse solution is obtained. Several sparse algorithms have been developed to obtain reduced support vectors to improve the generalization performance of LSSVM. However, all of them iteratively look for support vectors in training datasets, which may are not the most superior choice for building the function model. In this paper, we propose a method of reconstructed support vectors based on the training datasets. The support vectors reconstructed are near the hyper plane of target function and uniformly distributed, which have more contribution to target function. In addition, the method we proposed converges at a faster rate than those iterative algorithms, because one-step selecting strategy is adapted without repeated training. To show the efficacy and feasibility of our proposed algorithm, some comparing experiments are conducted, which are all favorable for our viewpoints. That is, the method we proposed needs less number of support vectors to reach the almost same generalization performance, most important, which has the better robustness and accuracy prediction for the real operating mode. |
|
Keywords:density clustering; lssvm; reconstruct support vectors |
|