结合投影与近邻操作的支持向量快速筛选方法

Fast Method to Filter Support Vectors Combined with Operation of Projection and Nearest Neighbors' Selection

为减少支持向量机(SVM)的计算负担,提高运算效率,并保证分类精度,提出一种结合投影与近邻操作的支持向量快速筛选方法.该方法利用Fisher投影轴的全局特性将其作为SVM最优分类面的近似法方向,在该方向快速筛除大量非支持向量,将分类边界附近的样本集作为备选支持向量集,同时为解决投影操作未考虑样本局部结构信息造成的误删支持向量的问题,结合近邻操作回选样本空间中备选支持向量的近邻样本更新扩充备选支持向量集,以该子集中的样本作为SVM的输入.在多个UCI标准数据集上的实验结果表明,该方法在充分保证分类精度的前提下有效降低了SVM的计算负担,具有较好的推广性.

To reduce computational burden and improve operation efficiency of support vector machine （SVM） while ensuring classification accuracy, a fast method to filter support vectors combined with operation of pro- jection and nearest neighbors＇ selection was proposed. Considering the global characteristics of Fisher projec- tion, it could be viewed as the approximate normal directions of SVM optimal hyperplane and filtered out a large number of non-support-vectors in this direction. The samples near the classification obtained boundary were regarded as alternative support vectors set. Neighborhood operation was combined to solve the problem that some support vectors might be filtered out mistakenly regardless of the local structure information. A num- ber of nearest neighbors of the alternative support vectors were selected backward from the samples space to update and expand the alternative support vectors set. The sets was treated as the SVM input. The experimen- tal resuhs on several UCI standard data sets showed that the fast method had good generalization performance and reduced the computational burden effectively under the premise of fully guaranteed classification accuracy.