C-SVM在不同类别样本数目不均衡下的优化

Optimization of C-SVM in case of samples with unequal numbers in their different varieties

在解决故障检测等分类问题时,若不同类别样本数目相差很大,C-SVM训练的分类错误总偏向于样本数较少的类别,因而影响了分类的精确性.为提高精确性,提出一种优化算法,在训练过程中针对不同类样本,采用不同的权值来优化训练过程,按正负类样本在总样本中所占的比例,加大样本数较少的类别权值,降低样本数较大的类别权值来实现两类样本间的均衡.实验结果表明,该方法对两类样本数目相差很大的问题有效.

In solving the problem of trouble-locating in the case of samples with great difference in their number for their different varieties, the training with C-SVM was undesirably under bias towards those varieties with fewer samples, so that the training accuracy was unsatisfactory. In order to improve its accuracy,an optimization algorithm was proposed based on taking different weights for different classes in the process of training. According to the proportion of positive and negative samples in the total samples, the weight for the minor variety with fewer numbers of samples was increased and the other decreased, so that the balance between two samples varieties was realized. It was showed by experiments that the proposed approach could improve the accuracy of classification.