摘要
为了提高支持向量机在大规模数据集处理时的精度,提出了基于核空间和样本中心角度的支持向量机算法。在核特征空间下,求得原训练集的两类中心点和两个中心点的超法平面,并获取原训练集样本到超法平面距离和到两中心点中点的比值,用比值最小的n个样本点替代训练集。给出的数学模型显示,该算法不需要计算核空间,比现有的同类缩减策略保留了更多的支持向量数目。结合实例对算法进行了仿真实验,实验结果表明,与同类算法相比,该算法在基本没有降低训练速度的情况下获得了更准确的训练精度。
To improve the training accuracy of support vector machine when processing large-scale data sets,in kernel-induced feature spaces,a support vector machine algorithm based on kernel-induced feature spaces is proposed.Firstly the two centers of the original training sets and vertical hyperplane of the two centers are gotten,then the ratio of the distance from the sample in original training sets is obtained to the vertical hyperplane and distance from it to the midpoint of the two centers,finally n samples are used with the smallest ratios to train instead of original training sets.The last mathematical model shows that the algorithm does not require calculation of the kernel-induced feature spaces,can retain more incremental support vectors to ensure the training accuracy than the other decrement strategies.With examples,a simulation analysis of the algorithm is given out.The results show that,compared with similar algorithms,it gets more training accuracy without training speed reduced basically.
出处
《计算机工程与设计》
CSCD
北大核心
2012年第2期586-590,共5页
Computer Engineering and Design
基金
重庆市科委攻关项目森林健康监测系统基金项目(CSTC
2009AC2068)
关键词
支持向量机
大规模
核空间
超平面
样本中心
support vector machine
Arge-scale
kernel-induced feature space
hyperplane
samples＇ center
