An Outlier Detection Technique Based on Spectral Clustering

Abstract

Spectral clustering receives much attention as a competitive clustering algorithms emerging in recent years, which has achieved excellent efficiency. Outlier detection shows its increasingly high practical value in many application areas such as intrusion detection, fraud detection, discovery of criminal activities in electronic commerce and so on. In this paper, we proposed a new outlier detection method inspired by spectral clustering. Our algorithm combines k-Nearest Neighbor and statistical techniques to acquire and use the information of eigenvalues and eigenvectors of feature space and finally digs the abnormal data as outliers and have achieved the method to directly compute k-neighbors spectral clustering. In this way, it not only effectively reduces the storage cost required for clustering, but also provides important reference values in terms of dimension disaster on the high-dimensional space based on distance and density-based outlier detection. We compare the performance of our method with distance-based outlier detection methods and density-based outlier detection methods. Experimental results show that the algorithm has higher accuracy and better applicability in outlier detection.