Data-Driven RAN Slicing Mechanisms for 5G and Beyond Sihem Bakri * , Pantelis A. Frangoudis ‡ , Adlen Ksentini * , and Maha Bouaziz § * EURECOM, Sophia Antipolis, France ‡ Distributed Systems Group, TU Wien, Vienna, Austria § University of Strasbourg, France Email: * {name.surname}@eurecom.fr, ‡ pantelis.frangoudis@tuwien.ac.at, § maha.bouaziz@unistra.fr Abstract—One of the main challenges when it comes to deploying Network Slices is slicing the Radio Access Network (RAN). Indeed, managing RAN resources and sharing them among network slices is an increasingly difficult task, which needs to be properly designed. The goal is to improve network performance and introduce flexibility and greater utilization of network resources by accurately and dynamically provisioning the activated network slices with the appropriate amounts of resources to meet their diverse requirements. In this paper, we propose a data-driven RAN slicing mechanism based on a resource sharing algorithm running at the Slice Orchestrator (SO) level. This algorithm computes the necessary radio resources to be used by each deployed network slice. These resources are adjusted periodically based on current estimates of achievable throughput performance derived from channel quality informa- tion, and in particular from the Channel Quality Indicator (CQI) values of the users of each network slice retrieved from the RAN. CQI information is reported to base stations by the User Equipment (UE) following standard procedures, but extracting and frequently reporting it from base stations to the SO may result in significant communication overhead. To mitigate this overhead while maintaining at the SO level an accurate view of UE channel qualities, we propose a machine learning approach to infer the stability of UE channel conditions, as well as predictive schemes to reduce the CQI reporting intensity based on the inferred channel status. Through extensive simulations, we demonstrate the efficiency of our data-driven RAN slicing framework, which allows to meet the stringent requirements of two main classes of network slices in 5G, i.e., enhanced Mobile Broadband (eMBB) and Ultra-Reliable and Low-Latency Communication (URLLC). Index Terms—Network slicing, radio resource sharing, RAN monitoring, CQI, machine learning. I. I NTRODUCTION The new generation of mobile networks, known as 5G, is expected to be launched by the end of 2020, promising the support of several novel use-cases coming from other industry sectors (or vertical industries), such as Industry 4.0, autonomous driving, entertainment, etc., in addition to the classical broadband connectivity. Building on the concept of network virtualization, Network Slicing is considered as one of the main enablers of 5G. It allows sharing a common phys- ical infrastructure through building virtual network instances (network slices) tailored to meet specific service requirements. 5G supports three types of network slices [1], namely en- hanced Mobile Broadband (eMBB), Ultra-Reliable and Low- Latency Communication (URLLC), and massive IoT (MIoT). Each network slice covers a set of services having the same re- quirements in terms of Quality of Service (QoS). Generally, a network slice is described and composed by a set of virtual and physical resources, in the form of Network Functions (known as VNF and PNF), deployed and interconnected together on top of a shared infrastructure. Indeed, VNFs include Core Network functions and slice-owner’s network services, while a PNF is an already deployed entity, such as a base station (BS) component. The network slice’s VNFs and PNF are deployed over different technological domains: Radio Access Network (RAN) and Cloud domain (including the Edge). Mostly, VNFs are deployed on Cloud and Edge domains, while PNFs mainly pertain to the RAN, e.g., gNodeB (gNB) baseband units. All the running network slices are isolated from each other, even though they share the same physical infrastructure. Management and orchestration of network slices is a critical task, involving making slice life-cycle management decisions at run-time based on monitoring feedback from the service and infrastructure levels. Different slice components at the RAN, transport, and core network need to be coordinated in such a way that target key performance indicators (KPIs) for throughput, latency, availability, and other metrics are attained. The increase in the number of coexisting slices and the number of diverse slice components that need to be monitored already come with significant overhead and the strain on the slice management and control planes is only expected to grow in beyond-5G settings. While the 5G network is maturing, the discussion about how future generations will look like has been kicked off [2], [3]. The general consensus is that the 5G network management architecture needs to evolve to meet complexity and heterogeneity-related coordination challenges by dealing with complex monitoring, analysis, and decision making and becoming more intelligent by natively supporting AI-driven operation. This need is a consequence of the ex- pected increase in the dynamics of the network as a result of extreme device mobility, massive densification, and the