Coulomb’s law-inspired parameter-free outlier detection algorithm

Abstract

Outlier detection plays a crucial role in the field of data mining and machine learning. Despite the diverse range of outlier detection algorithms, most still face two main challenges: i) the difficulty of choosing neighborhood parameter and top-n, existing algorithms can only solve one of them; ii) the hardship of simultaneously detecting global outliers, local outliers, and outlier clusters. To sovle these shortcomings, we propose a novel outlier factor, called Local Coulomb Outlier Factor (LCOF), inspired by Coulomb’s law in physics. It can measure the outlier degree of data objects without parameters due to the use of coulomb force and natural neighbors. LCOF first applies natural neighbors to adaptively derive the neighborhood parameter k. It further employs the interquartile range method to identify the top-n data points with the largest LCOF values as outliers. Finally, it can identify global outliers, local outliers and outlier clusters simultaneously completely without parameter k and top-n. Extensive experiments demonstrate that LCOF achieves the highest average F1 and Acc scores, with a maximum value of 0.99, outperforming most existing algorithms.