Model selection for SVM based on similarity margin of inner product

By investigating the relationships among generalization performance, VC dimension and margin in SVM, this paper proposes a novel similarity margin of inner product. Then, the richness or flexibility of corresponding function class, along with estimative upper bound of relevant dimensionality, is analyzed in detail. Aiming at the demerit of difficult selection of optimal kernel function and its parameters in SVM, we discuss the improvement of generalization performance in terms of similarity margin of inner product and construct a model selection approach based on similarity margin of inner product, by which we could effectively avert such shortcomings as high computational costs and formulation complexity in traditional model selection methods. Experimental results on both artificial dataset and practical dataset show that our algorithm can find out preferable kernel parameter model, as well as maintain better classification accuracy, so it is more applicable.