Mitigating concept drifts in proactive mobile core scaling via online learning models

Abstract

Mobile networks are in constant evolution, and at each generation new use cases emerge that challenge the current technology. Current 5G and future 6G use cases require mobile networks to provide high reliability and low latency, thereby stressing the entire infrastructure, and particularly the mobile core. To meet these stringent requirements, current core network technology incorporates scalability techniques based on machine learning models, which allow the anticipation of traffic demands and the proactive adjustment of network resources before any degradation in service quality occurs. However, changes in user habits, applications, and network protocols can alter the statistical properties of mobile traffic, a phenomenon known as concept drift. As a result, conventional machine learning models, trained on historical data, tend to lose accuracy over time, which may hinder the adoption of use cases with strict performance requirements. To address this limitation, this study evaluates the use of online learning models for proactive scaling of network functions in the mobile core. This approach seeks to employ models that can continuously adapt to changes in traffic distributions, maintaining consistent predictions even in scenarios subject to concept drift. As a case study, this master’s thesis investigates how online learning models can enhance the scalability of the Access and Mobility Management Function (AMF) under different types of concept drift. Using real traffic data from a telecommunications operator, several concept drift scenarios were defined, and a comprehensive simulation of thirteen online learning strategies was carried out in a simulated mobile core. The results show how the accurate predictions of online models impact service metrics such as the number of lost requests and the utilization level of the function. Moreover, the study highlights that not all online models maintain satisfactory predictive accuracy under traffic concept drift, which reinforces the importance of evaluating different models.