王建国, 陈肖洁, 张文兴
(内蒙古科技大学机械工程学院，内蒙古包头014010)
关键词：支持向量回归机; 模型选择; 回归（函数）估计; 样本特征.
中图分类号：TP18 文献标识码：A
摘要：为了解决支持向量回归机中核函数的核参数选择问题，在深入分析高斯尺度空间理论的基础上，提出了一种基于样本特征维数的高斯核的核参数选择算法。首先，初始化核参数值为样本总特征个数的倒数，选定样本输出的预测值与真实值的均方误差为评估标准；然后，利用评估标准最小化的原则来搜寻最优的核参数值；最后，应用最优的核参数值于支持向量回归机的训练和预测。UCI数据集上的实验结果表明了，相较于传统的5折交叉验证、最小二乘和神经网络算法，所提出的基于样本特征维数的核参数优化算法能高效地实现核参数的选择，收敛速度快，在回归估计中具有较好的泛化性能和预测精度，验证了该优化方法的有效性。
Research on Kernel Parameter Optimization of Support Vector Regression
WANG Jianguo, CHEN Xiaojie, ZHANG Wenxing
(School of Mechanical Engineering, University of Inner Mongolia Science and Technology, Baotou Inner Mongolia China014010)
Key words：support vector regression; model selection; regression estimation; sample feature
Abstract：For the kernel function parameter value selection issue of support vector regression (SVR), a modeling selection algorithm of Gaussian kernel was proposed based on the dimension of sample features together with analyzing the Gaussian Scale Space Theory in depth. First, the reciprocal of the total number of sample features was initialized as the original kernel parameter, and the mean square error between the predicted outputs and the true outputs was selected as the evaluation criterion. Then, the optimal kernel parameter researched by minimizing the evaluation criterion was applied in the training and testing of SVR. Compared with the conventional support vector regression algorithm of 5-fold cross-validation, the linear least squares and BP neural network algorithms, experimental results on UCI datasets show that the proposed algorithm can select the optimal kernel parameter value effectively, achieving the advantages of fast convergence, better generalization ability and higher precision in regression estimation of the model, which demonstrates the effectiveness of kernel parameter optimization algorithm of SVR.
.