An Effective Method for Combining Kernels with Class Separability
Ting‐Hua Wang, Qiong Zhang, Hai Hui Xie, Jun Ting Chen · Applied Mechanics and Materials · 2012
We propose a simple but effective method to determine the kernel weights for convex combination of multiple kernels. The key property of the proposed method is that it adopts a class separability criterion as the evaluation function to measure the goodness of the individual kernel. Based on the principle of class separability, we assign a weight to each kernel that is proportional to the quality of the kernel. Experimental results on Image Segmentation data set show the proposed method can improve accuracy in comparison with that using a single kernel or uniformly-combined kernel.