一种个性化k近邻的离群点检测算法

Outlier Detection Algorithm with Personalized k-Nearest Neighbor

在基于近邻思想的离群点检测算法中,参数k的选择是无法避免的,而k值过大或者过小都会对检测效果产生很大的影响.因此,如何选择k值是近邻方法研究中的重要内容之一.本文提出一种个性化k近邻(Personalized k-Nearest Neighbor,PKNN)的离群点检测方法,其每一个数据点的近邻个数是由算法自动确定,而不需要人为指定.位于稠密区域的点具有更多邻居,而位于稀疏区域的点具有更少的邻居.因此,PKNN方法确定的个性化近邻参数,更符合数据集的直观分布.实验结果表明,与现有方法相比,PKNN算法有很好的离群点检测效果.

For the outlier detection algorithms based on nearest neighbor,it is inevitable to determine the value of parameter k,and the selection of k has a decisive impact on the performance of outlier detection.So how to select the k value is one of the key issues for nearest neighbor methods.This paper proposes an outlier detection method of personalized k-nearest neighbor(PKNN),while the number of neighbors of each sample is automatically determined by the algorithm but not manually assigned.Intuitively,samples in dense regions have more neighbors,and those in sparse regions will have fewer neighbors.Therefore,the personalized nearest neighbor parameter may be more suitable to the distribution of data sets.Experimental results demonstrate that the proposed PKNN has good performance in outlier detection comparing with some state-of-the-art algorithms.