Ensemble2: Anomaly Detection via EVT-Ensemble Framework for Seasonal KPIs in Communication Network

TL;DR

This paper introduces Ensemble2, an anomaly detection framework for seasonal KPIs in communication networks, combining ensemble learning with EVT-based thresholding and online optimization for improved accuracy and efficiency.

Contribution

The paper presents a novel EVT-ensemble framework that automatically adjusts thresholds and incorporates online learning for real-time KPI anomaly detection.

Findings

Effective on production datasets

Runs at ~10 points/sec on standard hardware

Outperforms traditional threshold-based methods

Abstract

KPI anomaly detection is one important function of network management system. Traditional methods either require prior knowledge or manually set thresholds. To overcome these shortcomings, we propose the Ensemble2 framework, which applies ensemble learning to improve exogenous capabilities. Meanwhile, automatically adjusts thresholds based on extreme value theory. The model is tested on production datasets to verify its effectiveness. We further optimize the model using online learning, and finally running at a speed of ~10 pts/s on an Intel i5 platform.