摘要
为了解决原空间中最小二乘支持向量机的解缺乏稀疏性的缺点,提出了Pruning法、MFCV法和IMFCV法并对BDFS法进行了修改和运用。对一个不含有奇异点的系统而言,Pruning法、BDFS法和MFCV法在一定程度上都能实现原空间中最小二乘支持向量机解的稀疏性。BDFS法无论是训练时间还是预测时间都比Pruning法短;和MFCV法比起来,虽然BDFS法的训练时间短,但比MFCV的预测时间长。对一个含有奇异点的系统而言,Pruning法几乎失去了效用;虽然BDFS和MFCV法的训练时间都比IMFCV法的训练时间短,但IMFCV法能成功抑制奇异点从而缩短预测时间。
In order to solve the shortcoming of lacking sparseness in the solution of least squares support vector machine in the primal, the pruning method, multi-fold cross validation (MFCV) method and improved multi-fold cross validation (IMFCV) method are proposed and the backward deletion feature selection (BDFS) method is modified and applied. These methods all realize the sparseness of the least squares support vector machine to a certain degree in the primal for a system without outliers. The predicted time of BDFS is shorter than the pruning's; although the training time of BDFS is shorter than MFCV's, its predicted time is longer. In the face of a system with outliers, the pruning almost loses effectiveness, and the training time of BDFS and MFCV are both shorter than IMFCV's, but IMFCV can oppress outliers and reduces the predicted time remarkably.
出处
《系统工程与电子技术》
EI
CSCD
北大核心
2009年第1期142-145,237,共5页
Systems Engineering and Electronics
基金
国家自然科学基金资助课题(50576033)
关键词
最小二乘支持向量机
PRUNING
BDFS
MFCV
least squares support vector machine
pruning
backward deletion feature selection
multi-fold cross validation