自训练多项式光滑的半监督支持向量机  

Self-training Polynomial Smooth Semi-supervised Support Vector Machines

作　　者：刘叶青[1,2] 刘三阳[1] 谷明涛 

机构地区：[1]西安电子科技大学数学科学系 [2]河南科技大学理学院 [3]解放军96251部队

出　　处：《系统仿真学报》2009年第18期5740-5743,共4页Journal of System Simulation

基　　金：国家自然科学基金(60574075)

摘　　要：为了处理自训练半监督支持向量机算法中每次循环都需要求解二次规划因此效率低的问题,采用直接求解支持向量机的原始优化问题,由此得到一个不光滑的无约束优化问题。将正号函数展开为无穷多项式级数,由此得到了一族光滑函数,用多项式光滑函数对无约束优化问题进行逼近,并用共轭梯度算法求解模型。在人工数据和UCI数据集上的实验结果显示,给出的算法效率高,能保证标记样本很少时的分类精度并且不因标记样本的增多而明显提高分类精度。In order to deal with the problem of low efficiency brought by solving the quadratic programming in every cycle of self-training semi-supervised support vector machines, the method of training support vector machine in the primal was used. Therefore a nonsmooth optimization problem without constraint was deduced. Plus function was transformed to an equivalent infinite series. Thus a family of smoothing functions was derived. The polynomial smoothing functions was used to approach the optimization problem without constraint and the conjugate gradient algorithm was used to solve the model. Experimental results on artificial and real data support that the proposed algorithm with higher efficiency can guarantee the accuracy when the percentage of labeled sample is very low and the accuracy does not improve obviously as the number of labeled data increasing.

关 键 词：半监督学习 支持向量机 光滑 分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]