最大化约束密度单类分类器

计算机科学 ›› 2014, Vol. 41 ›› Issue (2): 59-63.

最大化约束密度单类分类器

赵加敏,冯爱民,陈松灿,潘志松

南京航空航天大学计算机科学与技术学院南京210016;南京航空航天大学计算机科学与技术学院南京210016;南京航空航天大学计算机科学与技术学院南京210016;解放军理工大学指挥自动化学院南京210007

出版日期:2018-11-14 发布日期:2018-11-14
基金资助:
本文受国家自然科学基金重点项目(61035003)资助

Maximum Constrained Density One-class Classifier

ZHAO Jia-min,FENG Ai-min,CHEN Song-can and PAN Zhi-song

Online:2018-11-14 Published:2018-11-14

摘要/Abstract

摘要： 针对单类分类器设计中的密度方法,采用以任务为导向的设计思想,通过人为指定核密度估计的密度函数上界,增强了边界低密度区域数据敏感性,同时也有效降低了密度估计的计算复杂度。进一步最大化全体样本的核密度估计函数并采用线性规划,可快速得到相应的稀疏解,因而称之为最大化约束密度单类分类器(Maximum constrained density based one-class classifier,MCDOCC)。为充分利用单类数据中可能出现的极少量异常数据,进一步提出了带负类的最大化约束密度分类器(MCDOCC with negative data,NMCDOCC),通过挖掘异常数据的先验信息来修正仅有正常类的数据描述边界,可提高分类器泛化能力。UCI数据集上的实验结果表明,MCDOCC的泛化能力与单类支持向量机相当,NMCDOCC较之则有所提高,从而能够更高效地估计目标类数据概率密度。

关键词: 单类分类器,概率密度估计,最大化约束密度,先验信息中图法分类号TP391.4文献标识码A

Abstract: A novel One-Class Classifier (OCC) was proposed within the framework of probability density estimation called Maximum constrained density based OCC,MCDOCC．By constraining the upper bound of the kernel density estimators with the introduced parameter,MCDOCC is more sensitive in the low-density region located on boundary,alleviates the computation cost at the same time．Then,through maximizing the average constrained density of the target data,MCDOCC optimizes the object function with linear programming and the sparse solution can be reached finally．To further improve the generalization ability,two ways for MCDOCC with Negative data (NMCDOCC) were developed for full utilizing the prior knowledge existed in outliers．Experimental results on UCI data sets show that the generalization ability of MCDOCC is comparable with one-class support vector machines,but NMCDOCC is better than it.

Key words: One-class classifier,Probability density estimation,Maximum constrained density,Prior knowledge

赵加敏,冯爱民,陈松灿,潘志松. 最大化约束密度单类分类器[J]. 计算机科学, 2014, 41(2): 59-63. https://doi.org/

ZHAO Jia-min,FENG Ai-min,CHEN Song-can and PAN Zhi-song. Maximum Constrained Density One-class Classifier[J]. Computer Science, 2014, 41(2): 59-63. https://doi.org/

参考文献

[1] 陈斌,陈松灿,潘志松,等．异常检测综述[J]．山东大学学报,2009,9(6):13-23
[2] Bishop．Neural networks for pattern recognition[M]．London:Oxford University Press,1995
[3] Sain S R,Gray H L,Woodward W A,et al．Outlier detection from a mixture distribution when training data are unlabeled[J]．Bulletin of the Seismological Society of America ,1999,9(1):294-304
[4] Tarassenko L,Hayton P,Brady M．Novelty detection for the identification of masses in mammograms[C]∥Fourth International Conference on Artificial Neural Networks．1995．London:University of Cambridge,June 1995
[5] Duda R O,Hart P E,Stork D G．Pattern Classification(2nd Edition ed)[M]．New York:John Wiley & Sons,2001
[6] Yeung D Y,Chow C．Parzen-window network intrusion detectors [C]∥Proceedings in the 16th International Conference on Pattern Recognition (ICPR’02)．2003
[7] Schlkopf B,Platt J C,Shawe-Taylor J．Estimating the support of a high-dimensional distribution[J]．Neural Computation,2001,13(7):1443-1471
[8] Tax D,Duin R P．Support vector domain description[J]．Pattern Recognition Letters,1999,20(11/13):1191-1199
[9] Dolia A,et al．Kernel ellipsoidal Trimming[J]．Computational Statistics and Data Analysis,2007,52(1):309-324
[10] 冯爱民,陈松灿．基于核的单类分类器研究综述[J]．南京师范大学学报,2008,8(4):1-6
[11] Xu Miao,Rahimi A,Rao R P N．Complementary kernel densityestimation[J]．Pattern Recognition Letters,2012,33(10):1381-1387
[12] Meinicke P,Twellmann T,Ritter H．Maximum contrast classifi-ers [C]∥Proceedings of ICANN’ 02Proceedings of the International Conference on Artificial Neural Networks．London,UK:Notes In Computer Science (LNCS) Press,2002:745-750
[13] Tax D,Duin R P．Support vector domain description[J]．Machine Learning,2004,54:45-66
[14] http://homepage.tudelft.nl/n9d04/occ/index.html

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed