To reduce the size of samples in training and accelerate the learning speed of support vector machine on large
scale datasets, an approach for preselecting support vectors by convex hull vertex method is proposed. Based on the fact
that the superset of support vectors could be formed by the convex hull vertexes of a linearly separating dataset, the duality
principle is applied to transform the solving of convex hull vertexes into the feasibility deciding of linear programmings.
Hence, all convex hull vertexes are accessible. Nonlinear mapping function is constructed to generalize the approach to
nonlinearly separating datasets. Experimental results on synthetic datasets and benchmark datasets show the effectiveness of
the proposed method.