基于误差最小的SVM最优分类面修正

针对C-支持向量机(C-SVM,C-Support Vector Machine)中惩罚系数C可能导致最优分类面不合理的问题,提出基于误差最小的SVM最优分类面修正方法.通过调整正负类分类间隔的约束条件,求解使训练样本总误差最小的偏置系数,并兼顾与正负类误差之差的绝对值的平衡,得到误差最小的更优分类面.实验证明该修正方法与C-SVM及其它修正方法相比,具有较高的分类精度和较强的抗噪声与野值数据干扰能力.

Modification of SVM's optimal hyperplane based on minimal mistake

Since some value of error penalties C in C-support vector machine (C-SVM) may cause extreme and irrational optimal separating hyperplanes, a new modification of SVM’s optimal hyperplane was proposed. By modifying the distance restriction of separating hyperplane between positive and negative classes, the bias coefficient was calculated with minimal training samples’ total error, while the absolute value of the error difference between positive and negative classes was balanced considered, a better separating hyperplane with minimal mistake was obtained. The experimental results show that this algorithm has improved the classified precision and enhanced the ability of reducing the outliers and noises’ effect, compared to C-SVM and other modification algorithm.