Fusing local density and approximate distance for nonparametric outlier detection

Abstract

Outlier detection is an essential yet challenging task in intelligent data analysis, and some density-based unsupervised methods have been introduced to identify outliers in low-density regions. However, these methods still suffer from inaccurate density estimation and limited capability in detecting diverse types of outliers. In this study, we propose a nonparametric outlier detection method with the fusion of density and distance (POD-FDD). The proposed method employs adaptive kernel density estimation based on natural neighborhoods, which reduces the sensitivity to parameters in density estimation. Moreover, the optimistic and pessimistic densities are introduced to enhance the reliability of density estimation in the local neighborhood. In addition, approximate reachability distance information is integrated to improve the capability of identifying cluster outliers. Ultimately, a robust parametric-free outlier detection method is developed to detect different types of outliers. Extensive comparative experiments and statistical significance analysis on synthetic and public datasets demonstrate its superior performance, achieving an average improvement of 1.97 % in the AUC metric.