一种基于近邻边界的粒度支持向量机学习策略

Learning Strategy Based on Neighboring-boundary Granular Support Vector Machine

针对粒度支持向量机进行粒划分后提取代表点时丢失部分重要分类信息从而影响分类准确率的情况,提出了一种基于近邻边界的粒度支持向量机(Neighboring-boundary Granular Support Vector Machine,NGSVM)的学习策略。首先采用kmeans方法进行粒划分,对不同的粒依据不同的规则提取粒内代表点,并按照要求分别将代表点放入精简集或修正集中,再用这两个集合中的数据对支持向量机进行训练;形成分类器后,根据核空间距离提取靠近分类面的近邻边界数据以对分类面进行修正。仿真实验结果表明,通过提取靠近分类面的近邻边界数据进行重新训练,能够修正分类面,进一步提高粒度支持向量机的分类准确率。

Granular support vector machine will lead to loss of partial classification information and accuracy degradation while dividing granules and extracting representative points.To solve this problem,a learning strategy based on neighboring-boundary granular support vector machine （NGSVM） was proposed.Samples were divided into granules with kmeans method firstly,different granules were dealt with different rules to extract representative points,and then these representative points were put into fixed set or reduced set as requested,by which support vector machine （SVM） was trained.After completion of classifier,classification plane would be rectified by extracting neighboring-boundary samples according to the kernel distance.The simulation results show that NGSVM gains a higher classification accuracy by extracting neighboring-boundary samples near classification plane and fixing classification plane.