A comparative study of outlier detection algorithms

Abstract

Data Mining is the process of extracting interesting information from large sets of data. Outliers are defined as events that occur very infrequently. Detecting outliers before they escalate with potentially catastrophic consequences is very important for various real life applications such as in the field of fraud detection, network robustness analysis, and intrusion detection. This paper presents a comprehensive analysis of three outlier detection methods Extensible Markov Model (EMM), Local Outlier Factor (LOF) and LCS-Mine, where algorithm analysis shows the time complexity analysis and outlier detection accuracy. The experiments conducted with Ozone level Detection, IR video trajectories, and 1999 and 2000 DARPA DDoS datasets demonstrate that EMM outperforms both LOF and LSC-Mine in both time and outlier detection accuracy. © 2009 Springer Berlin Heidelberg.