Optimal regularization parameters selection for Laplacian support vector machine
Juntao Li, Jia Yingmin, Junping Du, LI Wen-lin · 2008
Laplacian support vector machine (LapSVM) is an attracting tool for semi-supervised classification with manifold regularization. In this paper, we devote to selecting the extrinsic and intrinsic regularization parameters. To this end, a fusion of training and validation levels is first proposed, based on which, the optimal regularization parameters selection problem can be cast as a standard semidefinite programming. Then, a hybrid manifold regularization algorithm is also developed, thus eliminating the difficulty of balancing between the ambient space and the intrinsic geometric of the data distribution. Finally, experiments are performed that verify the research results.