LSVDD：基于局部支持向量数据描述的稀有类分析算法  

LSVDD:Rare class analysis based on local support vector data description

作　　者：熊海涛[1] 吴俊杰[1] 刘鲁[1] 李明[2] 

机构地区：[1]北京航空航天大学经济管理学院,北京100191 [2]中国石油大学工商管理学院,北京102249

出　　处：《系统工程理论与实践》2012年第8期1784-1792,共9页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(70901002,90924020);高等学校博士学科点专项科研基金(200800060005,20091102120014)

摘　　要：在单类支持向量数据描述算法的基础上,提出了一种基于局部支持向量数据描述的稀有类分析算法:LSVDD,能够处理存在类重叠的类不平衡问题.该算法利用支持向量数据描述算法对各类样本分别进行单类学习,从而获得单类模型:然后对单类模型的概念重叠区域使用属性选择进一步进行局部单类学习,最后得到综合分类模型.在仿真数据集和UCI数据集上的实验结果表明,LSVDD能够有效和稳定地提高稀有类分析精度.As a hot topic in data mining society, rare class analysis （RCA） has been widely used in various application domains including financial fraud detection, network intrusion detection, facility failure diagnosis, etc. However, it is not until recently that researchers have realized the impact of complex data structures to the RCA problem. We propose a local support vector data description algorithm LSVDD for RCA based on SVDD, which has the ability to handle class imbalance problem with the presence of class overlaps. Specifically, LSVDD firstly uses SVDD to get one-class classification model for each class and finds the concept overlapping regions between different classes. Then, the regions are locally trained using SVDD again after attribute selections. Finally, the models for non-overlapping and overlapping regions are combined to form a complete RCA model. Experimental results on artificial and real-world UCI data sets demonstrate that LSVDD can improve the performances of RCA stably and effectively.

关 键 词：数据挖掘 稀有类分析 支持向量数据描述 属性选择 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]