摘要
为了求解一类支持向量机(1-SVM)的二次规划问题(QPP),利用该QPP的稀疏解集性质,提出了基于作用集的1-SVM递推式训练算法.将支持向量集设定为作用集,迭代地局部优化作用集以获得全局最优解,并引进递推式算法降低计算复杂度。不同于序贯最小优化(SMO)收敛目标函数的思路,该算法寻找支持向量在最优状态下的分布,对Karush-Kuhn-Tucker(KKT)条件不敏感,并可获得解析的最优值。仿真结果表明,本算法在计算时间和精度上均优于SMO,可有效地应用于1-SVM的大样本学习。
A recursive algorithm based on active set method was proposed considering the sparsity of the solution of quadratic programming problem （QPP） to solve the QPP of one-class support vector machine （1-SVM）. The global solution was obtained by the iteration of the local optimizations of active set with support vectors treated as the active set. And a recursive method was introduced to reduce the computational cost. The algorithm searched for the optimal distribution of support vectors unlike the sequential minimal optimization （SMO） algorithm focusing on the convergence of object function. Then the absolutely analytical solution was obtained without the sensitivity of Karush-Kuhn-Tucker（KKT） tolerance. The simulation results show that the algorithm outperforms SMO in time consumption and accuracy and can be efficiently applied to 1-SVM for large scale systems.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2009年第1期42-46,共5页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(60702023)
浙江省自然科学基金资助项目(Y107440)
关键词
一类支持向量机
作用集法
二次规划问题
序贯最小优化
one-class support vector machine （1-SVM）
active set method
quadratic programming problem （QPP）
sequential minimal optimization （SMO）
