摘要
针对最小二乘支持向量回归机的解缺乏稀疏性、预测速度慢等问题,采用向量相关分析在高维特征空间约简支持向量.为使约简模型能最佳逼近原模型,提出原模型与约简模型预测训练样本的平方误差和作为新性能评价准则.为得到最优约简模型,定义了离散加法、减法和乘法算子,并将新性能评价准则作为适应度函数,采用整数编码的差分进化算法进行全局优化.4个标准数据集实验结果表明,与前人提出的3种性能评价准则相比,新算法得到的约简模型具有更好的泛化性能,并且在泛化性能略有下降情况下,支持向量数目大幅减少.
Aiming at lack of sparseness of the solutions of least squares support vector regression machine which leads to slow prediction speed and other problems,the vector correlation analysis was employed to reduce the support vectors in the high dimensional feature space.In order to make the reduced model best approximate the original one,sum squared prediction errors of training samples between the reduced model and original one were taken as the novel performance evaluation criterion.Discrete addition,subtraction and multiplication operator were defined and the novel performance evaluation criterion was used as fitness function.The best reduced model globally optimized by integer coded differential evolution algorithm could be obtained.The experimental results on four benchmark datasets show that reduced model obtained by the novel algorithm has better generalization performance,compared with the other three performance evaluation criterions presented before.And reduced model obviously decreases support vectors at cost of little generalization performance.
出处
《哈尔滨工程大学学报》
EI
CAS
CSCD
北大核心
2011年第8期1012-1018,共7页
Journal of Harbin Engineering University
基金
国家自然科学基金资助项目(61074127)
关键词
最小二乘支持向量回归机
稀疏性
向量相关分析
差分进化
整数编码
支持向量约简
least squares support vector regression machine
sparseness
vector correlation analysis
differential evolution
integer coded
support vector reduction
