Data-Driven Estimation of Throughput Performance in Sliced Radio Access Networks via Supervised Learning

In 5G systems, Network Slicing (NS) feature allows to deploy several logical networks customized for specific verticals over a common physical infrastructure. To make the most of this feature, cellular operators need models reflecting cell and slice performance for re-dimensioning the Radio Access Network (RAN). For enhanced Mobility BroadBand (eMBB) services, throughput is regarded as a key performance metric since it strongly influences user experience. This work presents the first comprehensive analysis tackling cell and slice throughput estimation in the downlink of RAN-sliced networks through Supervised Learning (SL), based on information collected in the operations support system. Different well-known SL algorithms are tested in two NS scenarios with single-service or multi-service slices serving eMBB users. To this end, several synthetic datasets are generated with a system-level simulator emulating the activity of a sliced RAN. Results show that NS alters the correlation between network performance indicators and cell throughput compared to legacy RANs, thus being required a separate analysis for NS scenarios. Moreover, the best model to estimate throughput at cell/slice level may depend on the scenario (single-service vs multi-service slices). In all cases, the best models have shown an estimation error below 10 %.