摘要
支持向量分类器的两种分类模型是超平面和超球体,前者在有重叠类别的数据集上表现不佳,后者存在过适应问题.为此,本文提出了双效分类思想,在训练分类器过程中同时学习类间差异信息及类内特征信息,以克服上述问题并提高分类性能.进而,提出了具体实现算法,支持向量双效分类器(Doubled-Informed classifier based on Support vectors,DISV).DISV为各类生成收缩远离球,并基于此定义决策函数.收缩远离球的球面穿过类内密集分布区,并保持与其他类的最大远离.DISV辅以训练子集抽取策略和参数自适应调整策略以降低算法代价.实验表明,双效分类思想有效,其在心脏肥大数据集上的诊断结果优于同类算法.
Hyperplane and hypersphere are two models of classifiers based on support vectors. However,hyperplane-based classifiers behave moderately in the datasets with overlapping classes,and hypersphere-based classifiers are caught by over-fitting problem. For that,a double-informed classification idea is proposed in the paper. The idea learns the difference informati on among classes and the characters within individual class. In addition,the implemented algorithm,a Doubled-Informed classifier based on Support vectors（ DISV）,is presented. DISV constructs the shrunk distant hypersphere for each class and derives the decision function from the information of shrunk distant hypersphere. Shrunk distant hypersphere keeps furthest away from other classes,and the surface of the hypersphere crosses the dense regions of class. To reduce computation cost,DISV is equipped with the selection method of training data and the self-tuning method of parameters. Empirical evidence verifies the validation of double-informed classification idea,and the application to atrial hypertrophy diagnosis demonstrates the advantage of DISV over the peer classifiers.
作者
凌萍
荣祥胜
李雪
LING Ping;RONG Xiang-sheng;LI Xue(College of Computer Science and Technology, Jiangsu Normal University, Xuzhou 221116, China;Department of Training,Air Force Logistics of P. L. A,Xuzhou 221000,China;College of Information Technology, University of Queensland, Brisbane 4067, Australia)
出处
《小型微型计算机系统》
CSCD
北大核心
2018年第5期1113-1120,共8页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61105129
61502198)资助
江苏师范大学自然科学基金重点项目(15XLA07)资助
关键词
双效分类思想
支持向量
收缩远离球
数据抽取策略
参数自适应调整
double-informed classification idea
support vectors
shrunk ＆ distant hypersphere
data selection method
self-tuning method of parameters
