带虚警抑制的基于归一化残差的野值检测方法

Normalized Residual-based Outlier Detection with False-alarm Probability Controlling

野值检测,或称异常值检测是模式识别和知识发现中一个重要的问题。以往的野值检测方法难以有效地抑制虚警概率,针对这一问题,该文提出一种带监督情形下基于归一化残差(Normalized Residual,NR)的野值检测方法。首先利用训练样本计算待考查模式的NR值,其次比较NR值与野值检测门限的相对大小,从而判断待考查模式是否为野值。该文理论上推导了野值门限与虚警概率之间的关系表达式,以此为依据设置检测门限,可实现在少量训练样本情况下仍能抑制虚警率的目的。计算机仿真和实测数据测试验证了所提方法在野值检测和虚警抑制方面的优越性能。

Outlier detection, also called anomaly detection, is an important issue in pattern recognition and knowledge discovery. Previous outlier detection methods can not effectively control the false-alarm probability. To solve the problem, a supervised method based on Normalized Residual（NR） is proposed. Using the training patterns, it first calculates the NR value of the query pattern, which is compared with a predefined detection threshold to determine whether the pattern is an outlier. In this paper, the relationship between the threshold and false-alarm probability is theoretically derived, based on which an appropriate threshold can be chosen. In this way,the desired false-alarm probability can be obtained even when few training patterns are available. Simulations and measured data experiments validate the superior performance of the proposed method on outlier detection and false-alarm probability controlling.