How to Select Optimal Parameters for Neighborhood-Based Outlier Detectors?

Neighborhood-based outlier detectors play a vital role in outlier detection, a cornerstone in data science. However, these detectors often rely on their parameters, and finding the optimal values for these parameters can be challenging. To address this issue, we propose a novel approach called K and T Finder using neighborhood consistency (KTF). In KTF, k represents the number of nearest neighbors, and t signifies the threshold value for outlier score thresholding. The core concept behind KTF is rooted in the idea that normal objects should exhibit consistent outlier scores with their neighbors, while outlier objects should display inconsistent outlier scores. Unlike previous approaches where k and t are determined independently, KTF takes a unique approach by simultaneously identifying both parameters. This method computes a consistency value for each combination of k and t, and the optimal values of k and t are chosen by maximizing this consistency value. The experimental results show that the proposed KTF outperforms existing baseline methods.