Abstract:

Support vector machine algorithm becomes another important technique after neural networks in the field of machine learning, but the evaluation and choice for kernel function is not solved. Based on the structural risk theory, a quantity estimation is proposed though empirical risk and confidence interval, and an evaluation formula for kernel function is given. This article points out the default of traditional definition of empirical risk, and gives a new definition. The results of simulation experiment show the feasibility and effectiveness of the method.

Key words: kernel function, support vector machine, linearly separable degree, linearly dispersion degree, structural risk

摘要： 继神经网络方法之后，支持向量机成为机器学习领域中的有效方法，但是核函数的评价和选取问题一直存在。该文从结构风险出发，通过经验风险和置信区间2个方面对核函数的性能进行量化，给出评价核函数性能的公式，指出传统经验风险定义的缺陷，并提出了一个新的定义。实验证明了该算法的可行性和有效性。

关键词: 核函数, 支持向量机, 线性可分度, 线性密集度, 结构风险

CLC Number: 

• TP18