Knowledge based support vector machines
Toshihide Ibaraki, T. Fukunaga · International journal of engineering intelligent systems for electrical engineering and communications · 2005
Support vector machines (SVMs), especially nonlinear SVMs, are known to have high performance as classifiers of data. In this paper, we propose to construct a nonlinear SVM from a set of available prior knowledge on the problem domain and to determine their weights by using training data set, which we call the knowledge based SVM (KSVM). A basic tool for KSVM is the reduced SVM (RSVM) proposed by Y. -J. Lee and 0. L. Mangasarian, in which kernel functions represent such knowledge. A KSVM has an advantage that its behavior is highly understandable as we can see how the kernels representing prior knowledge are combined into a classifier. It is confirmed by computational experiments that KSVMs can have high performance. We also discuss the separability condition and theVC dimension of KSVM.