Abstract:

Aiming at the problem that Minimum Class Variance Support Vector Machines(MCVSVM) which utilize only information in the non-null space of the within-class scatter matrix in small sample size case, this paper presents a novel algorithm called Least-Square-based Minimum Class Variance Support Vector Machines(LS-MCVSVM). The optimization problem of LS-MCVSVM can be solved by using Newton optimization, and the small sample problem can be avoided efficiently. Experimental results on several real datasets show that LS-MCVSVM can improve the generating ability and reduce the training time greatly.

Key words: supervised learning, Minimum Class Variance Support Vector Machines(MCVSVM), optimization algorithm

摘要：

针对最小类方差支撑向量机(MCVSVM)在小样本情况下仅利用类内散度矩阵非零空间中信息的问题，提出基于最小二乘的最小类方差支撑向量机(LS-MCVSVM)算法，通过牛顿优化法迭代求解LS-MCVSVM的优化问题，从而有效解决了小样本问题。实验结果表明，相对于MCVSVM，LS-MCVSVM算法可进一步提高泛化能力，减少训练时间开销。

关键词: 监督学, 最小类方差支撑向量, 优化算法

CLC Number: 

• TP391