The emergence of cutting-edge technologies and services such as Extended Reality (XR) promises to change how people approach everyday living. At the same time, the emergence of modern and decentralized architectural approaches has ushered in a new generation of mobile networks, such as 5G, as well as outlining the roadmap for B5G and even beyond. These networks are expected to be the enablers for the realization of the metaverse and other futuristic services. In this context, quantifying the service performance is a key enabler for dynamic, environment-adaptive, and proactive network management. This work presents an ML-based (Machine Learning) framework that uses data from the network, such as radio measurements, statistics, and configuration parameters to infer the best ML models that fit diverse XR Key Quality Indicators (KQIs). The output models integrate feature engineering techniques that enhance model size and performance. The proposed framework comprises data preprocessing, model definition, training, tuning, and validation. Additionally, to select the best combination algorithm this work introduces a metric called PET score, which evaluates algorithm candidates in terms of error performance and prediction time. These are considerations that are needed for time-sensitive services like XR's. To validate our proposal, the 360-video service has been chosen to demonstrate the potential of this ML framework with a real XR use case. Likewise, this work aims to serve as a baseline for future research based on the E2E Quality of Experience (QoE)-based network management in conjunction with other enabler technologies, such as network slicing, virtualization, and MEC (Multi-access Edge Computing), among others.
