Sparse multiple kernel for least square support vector regression

Least squares support vector machine is a successful method for classification and regression. However, it lacks the sparseness. Combining the multiple kernel techniques with iterative strategy, we propose a sparse multiple kernel for least squares support vector regression. The learning task is completed in the empirical feature space. The kernel function can be automatically learned as a convex combination of empirical base kernels. Meanwhile, all the constraints generated by the training set are considered during the training process, and the sparseness can be arbitrarily predefined. The resultant solution can be obtained through solving a quadratic programming, a linear programming and a unconstrained optimization programming iteratively. Experiments on the UCI datasets demonstrate the feasibility and effectiveness of the proposed algorithm. Copyright © 2013 Binary Information Press.