# Adaptive Thresholding Heuristic for KPI Anomaly Detection

**Authors:** Ebenezer R. H. P. Isaac, Akshat Sharma  
**arXiv:** [2308.10504](https://arxiv.org/abs/2308.10504)  
**Date:** 2023-08-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.10504v1/

## Key points

- The paper addresses the issue of anomaly detection in time series data, specifically focusing on Key Performance Indicators (KPIs) in a business context.
- Existing anomaly detection solutions are often limited to specific outlier detectors, which restricts their applicability to broader anomaly detection use cases.
- The authors propose an Adaptive Thresholding Heuristic (ATH) that dynamically adjusts the detection threshold based on the local properties of the data distribution and adapts to changes in time series patterns.
- ATH considers both the expected periodicity and observed proportion of anomalies to derive the threshold, effectively capturing meaningful anomalies while filtering out irrelevant outliers.
- ATH can be used with any underlying seasonality decomposition method and an outlier detector that provides an outlier score.
- Experimental results using EON1-Cell-U dataset demonstrate that ATH is computationally efficient and scalable for near real-time anomaly detection.
- ATH is flexible when used with multiple forecasters and outlier detectors.

## Summary

The paper titled "Adaptive Thresholding Heuristic for KPI Anomaly Detection" addresses the issue of anomaly detection in time series data, specifically focusing on Key Performance Indicators (KPIs) in a business context. While numerous outlier detectors have been developed for time series data, not all outliers are necessarily anomalies of interest from a business perspective. Existing anomaly detection solutions are often limited to specific outlier detectors, which restricts their applicability to broader anomaly detection use cases. In the case of network KPIs, these indicators tend to exhibit stochastic behavior, resulting in statistical outliers that do not significantly impact business operations. Therefore, there is a need for a heuristic approach that captures the business definition of an anomaly for time series KPIs. To address this challenge, the authors propose an Adaptive Thresholding Heuristic (ATH) that dynamically adjusts the detection threshold based on the local properties of the data distribution and adapts to changes in time series patterns. The ATH heuristic derives the threshold by considering both the expected periodicity and the observed proportion of anomalies. By minimizing false positives and addressing concept drift, this approach effectively captures meaningful anomalies while filtering out irrelevant outliers. Importantly, ATH can be used in conjunction with any underlying seasonality decomposition method and an outlier detector that provides an outlier score. To validate their hypothesis, the authors tested ATH on EON1-Cell-U, a labeled KPI anomaly dataset produced by Ericsson. The experimental results demonstrate that ATH is computationally efficient, making it scalable for near real-time anomaly detection. Additionally, it proves flexible when used with multiple forecasters and outlier detectors. Overall, this paper introduces an innovative approach to KPI anomaly detection by proposing the Adaptive Thresholding Heuristic (ATH). By dynamically adjusting the detection threshold based on local data properties and adapting to changes in time series patterns while minimizing false positives , ATH effectively captures meaningful anomalies . The experimental results validate its computational efficiency and scalability for real-time anomaly detection , making it a valuable tool in various business contexts .