Coupling Variable Selection and Anomaly Detection: Record-Based Approach
Michel Kamel, Anis Hoayek, Mireille Batton‐Hubert · 2025
The rapid expansion of interconnected devices globally requires telecommunication operators to manage complex networks efficiently. Advanced systems are needed to assist network engineers in maintaining these networks. Devices, or network elements, continuously transmit key performance indicators (KPIs). Current anomaly detection methods face limitations, particularly in addressing high-dimensional data and capturing extreme events. This paper introduces a novel record-based approach that reduces dimensions by focusing on the behavior of the tails of variables, offering a more robust alternative to traditional methods. Additionally, we propose an anomaly scoring system based on records theory, which enables real-time anomaly detection with computational efficiency and provides insights for root cause analysis. The proposed method significantly advances the state-of-the-art by integrating dimension reduction tailored to extreme events and an innovative threshold selection methodology, demonstrated using a real-world telecommunication dataset.