A Reformative KMSE Algorithm Based on the Numerical Analysis of Matrix

In order to improve the feature extraction efficiency of KMSE, we propose a novel KMSE algorithm. This algorithm assumes that the discriminant vector can be expressed as the linear combination of “significant nodes”, a subset of the training samples rather than all training samples. Determining the “significant nodes” based on the numerical analysis of the kernel matrix is the key of this method. We evaluate the effect, on the condition number of the kernel matrix, of each training sample. The rationale is that the lower the condition number of the kernel matrix is, the higher the generalization ability of the KMSE solution is. Exploiting the determined “significant nodes”, we construct a reformative KMSE model that can perform feature extraction computationally more efficient than naïve KMSE.