摘要
为了处理半监督支持向量分类优化中的非凸非光滑问题,引入一族多项式光滑函数来逼近非凸的目标函数,给出的多项式函数在样本的高密度区逼近精度高,逼近精度低时出现在样本的低密度区,同时可以根据不同的精度要求选择不同的逼近函数。采用BFGS算法求解模型。在人工数据和UCI数据集上的实验结果显示,算法不仅能保证标号数据很少时的分类精度,而且不因标号数据的增多而明显提高分类性能,因此给出的分类器性能是稳定的。
In order to solve the nonconvex and nonsmooth problem of semi-supervised support vector classification, a series of polynomial smooth functions were introduced in this paper which were used to approach the nonconvex objective function. The introduced polynomial functions have a high approximation accuracy in high density regions of samples and poor approximation performance appear in low density regions of samples. Different polynomial functions can be used according to the corresponding accuracy demand. The model was solved by the BFGS algorithm. Experimental results on artificial and real data support that the proposed algorithm can guarantee the accuracy when the percentage of labeled sampled is very low and the accuracy did not improved obviously as the number of labeled data increasing. The performance of the proposed classifier is stable.
出处
《系统工程理论与实践》
EI
CSCD
北大核心
2009年第7期113-118,共6页
Systems Engineering-Theory & Practice
基金
国家自然科学基金(60574075)
关键词
半监督学习
支持向量机
分类
模式识别
semi-supervised learning
support vector machine （SVM）
classification
pattern recognition
