PRUNE SUPPORT VECTOR MACHINES BY AN ITERATIVE PROCESS

摘要

Support vector machines (SVMs) are well known to give good results on many pattern recognition problems, but they exhibit classification speeds that are substantially slower than those of neural networks. One can speed up SVM classification by reducing the complexity of the decision function, which can be obtained by decreasing the number of support vectors. An iterative process is proposed to prune SVM and avoid obvious decline in classification accuracy. Computational results indicate that the number of support vectors is related to the size of training set, and that simplified SVMs with many fewer support vectors have classification accuracy nearly equal to that of original SVM. The proposed method is also compared with previous research, with results that support it as an effective method to obtain a simplified SVM for large problems.