A Relative Kernel-density Based Outlier Detection Algorithm

Abstract

An outlier is a point that could impact the data quality and the analysis result of data mining. The outlier detection may also be viewed as the pre-processing step for finding the objects that do not ensue the well-defined notions of predicted behavior in a data set. Recently, it emerges as a challenging issue in the real-world scenario to discover rare or novel events. This paper presents a novel and effective outlier detection method with kernel density estimation (KDE). A relative kernel-density based outlier factor (KDOF) is introduced to measure the outlier-ness score of an object in a given data sets. The proposed method is categorized into three phases. The first phase is to compute the local density at the given point using the KDE procedure. Then density fluctuation (DF) of each data point is calculated in step with the sum of the squares of the density differences between the point and its neighbors. Finally, the KDOF for each point is computed by comparing the density fluctuation at a given point with the average density fluctuation of its neighbors. Our experiments carried out on five real data sets have proven that the proposed method can perform better over other state-of-the-art outlier detection methods.