An efficient local outlier detection approach using kernel density estimation

Abstract

In recent times, outlier detection has played a crucial role in computer networks, fraud detection, and many such applications. Despite adequate research initiatives addressing the topic of finding outliers in datasets, still faces numerous obstacles in establishing an appropriate approach for addressing specific applications of interest. The paper introduces an unsupervised outlier detection method, achieving robust local density estimation through the customization of a nonparametric kernel density evaluation. The identification of outliers involves comparing the local density of each data point with that of its neighbors. Additionally, the proposed method addresses the challenge of manually selecting the parameter for the size of the nearest neighborhood by assigning a predefined value to this parameter. With this predefined value of the parameter, the proposed method demonstrates efficient results, unlike other existing methods that require different values of this parameter for different datasets. To demonstrate the impact of this parameter and evaluate the performance of the proposed method, several assessments were done. The findings prove that the suggested method effectively detects local outliers.