基于子空间的可解释性多变量异常检测  

Subspace-based Multivariable Anomaly Detection with Interpretability

作　　者：宋润葵 郑扬飞[1] 郭红钰[1] 李倩 SONG Runkui;ZHENG Yangfei;GUO Hongyu;LI Qian(Department 8 of System,North China Institute of Computing Technology,Beijing 100083,China)

机构地区：[1]华北计算技术研究所系统八部,北京100083

出　　处：《计算机测量与控制》2022年第11期38-45,共8页Computer Measurement &Control

基　　金：科技创新2030—“新一代人工智能”重大项目(2020AAA0105100)。

摘　　要：为了解决在多维特征数据中,部分异常点被分散的特征空间所掩盖而无法检出的问题,以及缓解当前异常检测方法的结果可解释性差或不具有可解释性的状况,提出了基于子空间的可解释性多变量异常检测算法;首先在多维特征空间中,对每一个维度的特征进行分布检验,在此基础上为每一个对象选择出一个特征空间的集合,进而为每个对象计算出异常值分数;在此过程中,高效利用算法的中间过程产物,来对算法结果加以解释,改善使用者对数据的理解;使用真实数据集,对算法进行了验证,实验结果表明其具有较好的准确性和运行时间,并很好地解释了异常点的异常性。To solve the problem that some outliers cannot be detected,and the scattered feature space covers the outliers in the multidimensional feature data,which alleviates the phenomenon that the results of current anomaly detection methods are poorly interpretable or not interpretable,a subspace-based multivariate anomaly detection algorithm with interpretability is proposed.Firstly,in the multidimensional feature space,the distribution test of the features of each dimension is carried out.On this basis,a set of feature space is selected for each object,and then the outlier score is calculated for each object.In this process,the intermediate process products of the algorithm are efficiently used to interpret the algorithm results and improve users to understand the data.By using the real data set,the algorithm is verified.The experimental results show that it has good accuracy and running time,which well explains the abnormity of outliers.

关 键 词：异常检测 可解释性 点异常 统计型数据 机器学习 无监督学习 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]