A new method for optimizing the combinational kernels
Xue Tian, Xu Yang · 2010
The optimal kernel selection is a critical problem for the kernel-based learning algorithm. In order to obtain good results, the kernel function must be chosen in a data-dependent manner. To this end, we propose a new feature space based class separability measure to evaluate the conformation of kernels to the data. The optimal combination coefficients of multiple Gaussian functions are obtained by optimizing this measure. Experimental results show that our algorithm outperforms the cross-validation method and the radius margin bound method, and moreover, can further improve the performances of SVM classifiers.