一种改进的 LOF 异常点检测算法-《计算机技术与发展》

文章信息/Info

Author(s):: ZHOU Peng; CHENG Yan-yun; School of Automation,Nanjing University of Posts and Telecommunications,Nanjing 210023,China

摘要:: LOF 异常点检测算法在实际应用中有两个缺陷:一是离群因子值只与参数 K 有关,当 K 取值不同时,离群因子的值将不同,之前是异常点的数据可能不再是异常点。二是对于未知异常点个数的数据集,选择参数 K 以保证离群点的挖掘数量合理难以做到。因此,提出一种结合平均密度的改进 LOF 异常点检测算法。首先分析数据集中数据点的平均密度,根据密度的分布情况确定数据集的异常点个数 M 1 及异常集 D 1 ,然后通过计算离群因子确定 M 2 ( M 2 = M 1 )个异常点及异常集 D 2 。取 D 1 与 D 2 的交集作为最终的离群集。实验结果表明,改进算法在检测精准性方面有显著提高,误报率较低,综合评价指标 F 值比 LOF 算法有显著增强。

Abstract:: In practical application,LOF,an anomaly detection algorithm,has two defects. One is the outlier factor value only related to the parameter K . When K is changed,the value will be different from before and an abnormal point may be a normal point. Another is for a data set with unknown abnormal points. It is very hard to choose a suitable parameter K to ensure reasonable mining number of outlier points. Therefore,an improved LOF combined with the average density is proposed. Firstly,the average density of each point is analyzed, and the number of abnormal points ( M 1 ) and abnormal set ( D 1 ) are determined according to the distribution of average density in the data set. Then M 2 ( M 2 = M 1 ),another number of abnormal points,and D 2 ,another abnormal set,are ensured through calculating the value of outlier factor. The intersection of D 1 and D 2 is taken as the final result. Experiment shows that the improved algorithm can improve the detection precision remarkably with lower false rate,and is superior to LOF on the comprehensive evaluation index F .