Assessing Anomaly Detection Algorithms in Mobile Networks

Abstract

Accurate and timely anomaly detection is critical for ensuring the reliable performance of modern telecommunication networks. However, monitoring vast amounts of operational data in real-time poses challenges for both human experts and traditional rule-based detection systems. This paper empirically evaluates several contemporary machine learning-based algorithms using a real-world dataset capturing metrics from EUtranCell. Models are assessed based on their detection of anomalies previously identified by domain experts. Additionally, this paper conducts an in-depth analysis of total cases flagged to evaluate calibration beyond reported matches. Results show that while some models detect most known anomalies, histogram examination reveals these models tend to label numerous observations near operating thresholds as outliers. Overall, the study examines the capabilities and limitations of different anomaly detection techniques for supporting cognitive telecom network operations.