基于局部线性重构的主动支持向量机分类器构建方法  

Construction Method of Active Support Vector Machine Classifier Based on Local Linear Reconstruction

作　　者：宋丹[1,2] 石勇[3] 邓宸伟[4] 

机构地区：[1]中国科学院大学博士后科研流动站,北京100190 [2]中国工商银行博士后科研工作站,北京100032 [3]中国科学院大学管理学院,北京100190 [4]北京理工大学信息与电子学院,北京100081

出　　处：《系统工程》2016年第10期135-141,共7页Systems Engineering

基　　金：国家863项目(2015AA8017032);国家自然科学基金资助项目(61301090)

摘　　要：针对基于距离的主动支持向量机(Distance-Based Active SVM,DASVM)在选取初始标记样本时采用了随机的策略而影响了分类器学习效率和分类性能的问题,提出了一种基于局部线性重构的主动支持向量机分类器构建方法(CLASVM)。CLASVM在DASVM分类器构建方法的基础上增加了基于类内-类间局部线性重构的训练样本初选策略,该训练样本初选策略以"聚类假设"和"流形假设"的思想为指导,选取若干同时具有最大类间重构误差和最小类内重构误差的样本作为初选样本进行标记。实验表明,CLASVM在样本初选阶段能够选取较多的支持向量,样本标记成本减少,训练效率与分类器性能与其他参与对比的分类器相比有显著提升。To solve the problem of DASVM Classifier which uses the random strategy to select initial labeled samples, this paper proposes an active SVM classifier construction method based on local linear reconstruction （CLASVM）. On the basis of the construction method of DASVM classifier, CLASVM adds training sample primary strategy based on intra- class and extra-class local linear reconstruction. This training sample selection strategy is guided with cluster assumption and manifold assumption and selects a number of samples which has the biggest extra-class reconstruction error and the smallest intra-class reconstruction at the same time. Experimental results slaow that CLASVM selects a large number of support vectors in the primary stage, and the cost of sample labeling is reduced, the training efficiency and classifier performance are significantly improved compared with other classifiers.

关 键 词：样本初选 主动支持向量机 局部线性重构 支持向量 主动学习 

分 类 号：TP180[自动化与计算机技术—控制理论与控制工程]