[1]程学云,吉根林,凌霄汉.基于SVM的多类代价敏感学习及其应用[J].南京师范大学学报(工程技术版),2006,06(04):079-82.
　CHENG Xueyun,J I Ge n lin,et al.SVM-Based Multiclass Cost-Sensitive Learning and its Application[J].Journal of Nanjing Normal University(Engineering and Technology),2006,06(04):079-82.
点击复制

基于SVM的多类代价敏感学习及其应用
分享到：

南京师范大学学报（工程技术版）[ISSN:1006-6977/CN:61-1281/TN]

卷:
06卷
期数:
2006年04期
页码:
079-82
栏目:
出版日期:
2006-12-30

文章信息/Info

Title:
SVM-Based Multiclass Cost-Sensitive Learning and its Application
作者:
程学云1;  2 ; 吉根林1 ; 凌霄汉1
1. 南京师范大学数学与计算机科学学院,江苏南京210097; 2. 南通大学计算机科学与技术学院,江苏南通226007
Author(s):
CHENG Xueyun1;  2 ;  J I Ge n lin1 ;  L IN Xiao han1
1. School ofMathematics and Computer Science, Nanjing Normal University, Nanjing 210097, China;
2. School of Computer Science and Technology, Nantong University, Nantong 226007, China
关键词:
代价敏感学习;  支持向量机;  入侵检测;  漏报率;  误报率
Keywords:
cost-sensitive learning;  support vectormachine ( SVM) ;  intrusion detection;  false negative;  false posi-tive
分类号:
TP181
摘要:
标准的分类器设计一般基于最小化错误率.在入侵检测等问题中,不同类型的错分往往具有不等的代价.通过在支持向量机的类概率输出中引入代价敏感机制,提出了3种基于最小化总体错分代价设计分类器的方法.实验结果表明通过改变代价矩阵,能在漏报率、误报率及稀有类样本的错误率之间调节,从而保证在误报率尽可能小的情况下降低漏报率和稀有类样本的错误率,以减少总体错分代价.
Abstract:
The standard classifier is usually based on minimizing the error rate, but in intrusion detection and some p ractical p roblems, different errors have different costs. Three kinds of support vector machine ( SVM) learning methods based on minimizing the totalmisclassification cost are p roposed, which introduce the cost-sensitive mecha2 nism into the p robabilistic outputs of SVM. The results show thatwe can trade off among false negative , false positive and error rate of rare class by changing cost matrix, which can minimize false negatives and error rate of rare class while constraining false positives at a low level so as to minimize the totalmisclassification cost.

参考文献/References:

[ 1 ] MARK A DAVENPORT. The 2nu - SVM: A Cost2Sensitive Extension of the nu - SVM [ R ]. Rice University ECE Technical Report TREE 0504, 2005.
[ 2 ] CHANG Chihchung, L IN Chihjen. L IBSVM: A library for support vectormachines [ EB /OL ]. http: / /www. csie. ntu. edu. tw/ ～cjlin / libsvm, 2005.
[ 3 ] DOM INGOS P. MetaCost: a generalmethod formaking classifiers cost2sensitive[C ] / / Proc of the 5 th International Conference on Konwledge Discovery and DataMining. San Diego: ACM Press, 1999. 155-164.
[ 4 ] OSUNA E, FREUND R, GIROSI F. Support vectormachines: Training and applications[R ]. A IMemo 1602,MITA ILab, 1997.
[ 5 ] WU Tingfan, L IN Chihjen, RUBY CWeng. Probability estimates formulti2class classification by pairwise coup ling[ J ]. Journal ofMaching Learning Research, 2004 (5) : 975-1 005.
[ 6 ] KDD Cup 1999 Data [DB /OL ]. [ 1999 - 10 - 28 ] http: / /kdd. ics. uci. edu /databases/kddcup99 /kddcup99. html.
[ 7 ] SCOTT C, NOWAK R. A neyman2pearson app roach to statistical learning [ J ]. IEEE Transactions on Information Theory, 2005 (51) : 3 806-3 819.

相似文献/References:

[1]文益民,廖洪元,周立华.一种可减少训练时间的分层并行支持向量机方法[J].南京师范大学学报(工程技术版),2005,05(01):008.
　WEN Yimin,LIAO Hongyuan,ZHOU Lihua.A Hierarchical and Parallel Support Vector Machines Algorithm for Reducing the Training Time[J].Journal of Nanjing Normal University(Engineering and Technology),2005,05(04):008.

备注/Memo

备注/Memo:
基金项目: 江苏省自然科学基金资助项目(BK2005135)和南通大学校级自然科学研究基金资助项目(05Z053) .
作者简介: 程学云(19782) ,女,助教,硕士研究生,主要从事数据挖掘、模式识别的教学与研究. E-mail: chen. xy@ntu. edu. cn
通讯联系人:吉根林(19642) ,教授,博士生导师,主要从事数据库、数据挖掘与入侵检测技术的教学与研究. E-mail: glji@njnu. edu. cn
更新日期/Last Update: 2013-04-29