基于支持向量机的增量式算法  

Incremental learning algorithm based on support vector machine

作　　者：黄启春[1] 刘仰光[2] 何钦铭[1,2] 

机构地区：[1]浙江大学计算机科学与技术学院,浙江杭州310027 [2]浙江大学宁波理工学院,浙江宁波315100

出　　处：《浙江大学学报（工学版）》2008年第12期2121-2126,共6页Journal of Zhejiang University：Engineering Science

基　　金：宁波市自然科学基金资助项目(2008A610026);浙江大学宁波理工学院青年基金资助项目(1141147G601)

摘　　要：为了扩展支持向量机在大规模数据集和成批出现数据领域的应用,提出了一种基于支持向量机的增量式学习算法.利用标准的支持向量机算法训练得到初始的目标概念,通过增量式步骤不断更新初始的目标概念.更新模型是求解一个与标准支持向量机具有类似的数学形式的凸二次规划问题.证明了在可分情况下,如果新增加的样本不是位于边界区,那么增量式过程既不会改变分类平面也不会改变分类平面的表达.与现有的增量式支持向量机算法相比,该算法无需额外计算就可实现增量式的逆过程并且训练时间与增量式步骤数成反比.实验结果表明,该算法满足稳定性、能够不断改进性能以及性能回复三个准则.An incremental learning algorithm based on support vector machine was proposed to process large scale data or data generated in batches. Initial goal concept learned by standard support vector machine algorithm was updated by an updating model. The model solved a convex quadratic programming similar to standard support vector machine algorithm. The algorithm proves that not only classification hyperplane but also the representation of classification hyperplane arenot changed during incremental learning procedure if the new samples arenot in the boundary in separable ease. Decreasing learning procedure is easy to implement without extra computation and learning time is inversely proportional to incremental learning step compared with existing incremental learning support vector algorithms. Results show that the algorithm satisfies the three criteria of stability, improvement and recoverability.

关 键 词：机器学习 模式分类 支持向量机 增量式算法 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]