Coupling Variable Selection and Anomaly Detection: Record-Based Approach

The proliferation of interconnected devices is rapidly expanding globally, and, as a result, telecommunication operators are responsible for managing intricate and expansive networks. Consequently, there is a need for advanced and efficient systems to aid network engineers in maintaining these networks. Devices, which can also be referred to as network elements, continuously transmit essential performance data known as key performance indicators. By utilizing data derived from these metrics and implementing intelligent anomaly detection models, the devices can assist in determining the optimal production maintenance schedule for the network. As anomaly detection models deal with extreme events, this study proposes a method of reducing dimensions by focusing on the behavior of the tails of underlying variables, rather than the entire distribution. In addition to that, an anomaly scoring system, also based on records theory, is proposed, which has several advantages over current state-of-the-art models. The effectiveness of this approach is demonstrated by implementing it on a real-world dataset.