Efficient geometric algorithms for support vector machine classifier

The geometric approaches based on reduced convex hull (RCH) are promising methods for solving support vector machine (SVM), which have been the focus of intense theoretical as well as application-oriented research in machine learning. In this paper, two efficient geometric learning algorithms for SVM, termed as DNP-GA and PDNP-GA, are proposed by introducing the direct nearest point-pair and probabilistic speed-up strategies. Extensive experiments on several artificial and benchmark databases have been conducted to show that, compared with the corresponding geometric algorithm, the proposed algorithms degrade many kernel evaluations without loss of generalization.