基于流形学习的离群点检测方法  

The research of detection of outliers based on manifold learning

作　　者：徐雪松[1] 宋东明[1] 张谞[1] 许满武[2] 刘凤玉[1] 

机构地区：[1]南京理工大学计算机科学与技术学院,南京210094 [2]南京大学计算机科学与技术系,南京210093

出　　处：《中国工程科学》2009年第2期82-87,共6页Strategic Study of CAE

摘　　要：为了提高高维数据集合离群数据挖掘效率,提出了一种基于流形学习的离群点检测算法。局部线性嵌入(locally linear embedding,LLE)算法是流形学习中有效的非线性降维方法,它的优势在于只定义唯一的参数,即邻域数。根据LLE算法的思想寻找样本数据的内在嵌入分布,并通过邻域数选取和降维后数据点之间的距离调整,提高了数据集中离群点发现效率,同时利用离群点权值判别式进行权值数据判定,根据权值的大小标识出数据集中的离群点,仿真实验的结果表明了该方法能够有效地发现高维数据集中的离群点。与此同时,该算法具有参数估计简单、参数影响不大等优点,该算法为离群点检测问题的机器学习提供了一条新的途径。The data dimensionality reduction is the main method that can enhance the outliers mining efficiency based on higher-dimension data set. The research of detection of outliers based on manifold learning is proposed after analyzing the advantages and disadvantages of the classical outlier mining algorithm in the paper. Local Linear Embedding algorithm（LLE） is an effective technique for nonlinear dimensionality reduction in manifold learning. Compared with other dimensionality reduction algorithms, the advantage of the local Linear Embedding algorithm is that it only defines unique parameter, i.e. number of nearest neighbours. With the idea of Local Linear Embedding, the algorithm can select optimal parameter and regulate the distance among data set after data dimensionality reduction, so as to improve efficiency of detection of outliers. The algorithm determines weighted values by discretion formula of weighted outliers. Through these weighted values, the experts can identify the outliers easily. Simulation results illustrate that this algorithm is very efficient. Moreover, our method has the advantage of simple parameter estimation and low parmneter sensitivity. Our method gives a new way for the solution of detection of outliers.

关 键 词：流形学习 离群点检测 高维数据 维数约减 离群数据 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]