Dynamic Combination of SVM Based on Optimal Kernel

As a cost function, fisher linear discriminant criterion can be used to optimize the kernel function. However, the dataset may not be linearly separable even after kernel transformation in many applications. So, SVMs that use the kernel function optimized by fisher criterion can not ensure the performance. Motivated by this issue, an ensemble algorithm was proposed. Firstly, partitioning a dataset into a number of subsets; Secondly, constructing a localized optimal kernel function based on Fisher criterion in every subset, and then training a SVM using this kernel function. For a testing sample, computing its K-nearest-neighbor, then, choosing SVMs depending on its neighbors to ensemble. Experimented on UCI multiclass datasets, comparing with three other multiclass methods, results demonstrate that the proposed method can achieve a high precision, especially, it has high speed for big and multiclass training set.