Outlier Detection Method based on Improved K-means Clustering Algorithm

Outlier detection is one of the important methods of data mining and analysis. The previous K-means outlier detection methods are easy to be misled by outliers in the clustering process and cannot detect outliers well. To solve this problem, we propose a method of outlier detection based on an improved K-means clustering algorithm. This method adaptively detects edge points (including outliers and edge points of other clusters) through the local threshold obtained from the distance distribution of the data points in the cluster to the centroid. The edge points are ignored when the centroid is updated to avoid the situation that the edge points mislead the centroid update. After completing the clustering, the farthest point in each cluster is put into the outlier cluster iteratively. When the average distance of the data in the cluster reaches stability or reaches the maximum number of iterations, the iteration ends. This method retains the ease of use of traditional K-means which accelerates the speed of clustering iteration and finds outliers accurately in data sets. Experimental results show that the algorithm can effectively detect outliers and have good results on multiple data sets.