Fusion of Domain Knowledge and Model Based on Support Vector Permutation Kernel Function
Hui Li · Chinese Journal of Computers · 2002
This paper presents one theory and method for revising a support vector kernel function, by means of which permutation information containing the invariance common sense is used in the process of SVM training. Compared with the traditional methods, the introduction of the permutation kernel function provides a theory foundation and methodology, which supports the fusion of knowledge and model. First, in terms of the conception of permutation, the invariance common sense about the structure of object is formalized, and the conceptions of the syngensis set and syngensis permutation are put forward; then, the permutation transformation matrix is solved, which expresses the disturbance of object pattern. Under the constraint of the classification invariance, the kernel function is revised using permutation transformation matrix. As a result, the SVM classifier based on permutation function is obtained. The experiment shows that the method in this paper is an effective one to improve the generalization performance of the SVM classifier with the permutation information. This paper also revises the sufficient condition needed by the classification invariance and proves it, which ensures the validity of the method.