An exploratory comparison of outlier detection methods in typical performance time series

Abstract

This study compares and analyzes three outlier detection methods applied to server CPU utilization time series – IQR, Z-Score, and DBSCAN. Based on 1,435 data points collected over a 1-minute interval, the core contribution of this study is a parameter sensitivity analysis. The study demonstrates that there is no universally applicable «best» method and provides deeper insights into selecting and configuring outlier detection strategies in practical performance monitoring systems, highlighting the critical impact of understanding parameter tuning.