基于最近邻距离差的改进孤立点检测算法  

Improved outlier detection algorithm based on difference between nearest neighbors distance

作　　者：侯晓晶[1] 王会青[1] 陈俊杰[1] 刘茜[2] 

机构地区：[1]太原理工大学计算机科学与技术学院,山西太原030024 [2]北京理工大学计算机学院,北京100081

出　　处：《计算机工程与设计》2013年第4期1265-1269,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(60970059;61170136);山西省自然基金项目(2011011015-4);山西省青年基金项目(2011021013-3);太原理工大学校青年基金项目(K201021)

摘　　要：k最近邻孤立点检测算法的检测结果受用户设置参数的影响较大,并且无法判定孤立点强弱,针对该缺陷,引入阈半径和密集度阈值,提出基于最近邻距离差的孤立点检测算法。通过在多个数据集上的实验表明,改进算法扩大了参数的设置范围,降低了参数对结果的影响,并能够有效检测出强孤立点,用户通过调整密集度阈值,可以判定孤立点强弱,改进算法增强了原算法的稳定性和灵活性。Results of k nearest neighbor outlier detection algorithm are affected by parameters set by users deeply and are unable to determine the strength. In order to eliminate this defect, the threshold radius and density threshold is introduced and improved outlier detection algorithm is presented based on difference between nearest neighbors distance. The experimental results of sev- eral data sets show that improved algorithm extends the span of parameters, reduces the impact of parameters on the results and can effectively detect strong outliers. By setting the intensity threshold, users can determine the strength of the outliers. The improved algorithm enhances the stability and flexibility of the original algorithm.

关 键 词：孤立点检测 最近邻距离差 参数设置 k最近邻 强孤立点 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]