Near-boundary Data Selection for Fast Support Vector Machines

Support Vector Machines(SVMs) have become more popular than other algorithms for pattern classification. The learning phase of a SVM involves exploring the subset of informative training examples (i.e. support vectors) that makes up a decision boundary. Those support vectors tend to lie close to the learned boundary. In view of nearest neighbor property, the neighbors of a support vector become more heterogeneous than those of a non-support vector. In this paper, we propose a data selection method that is based on the geometrical analysis of the relationship between nearest neighbors and boundary examples. With real-world problems, we evaluate the proposed data selection method in terms of generalization performance, data reduction rate, training time and the number of support vectors. The results show that the proposed method achieves a drastic reduction of both training data size and training time without significant impairment to generalization performance compared to the standard SVM.