Network performance prediction using graph neural networks: application to network slicing

ENG- This thesis addresses key challenges in the optimization of network slicing in Beyond 5G (B5G) networks, focusing on the use of Graph Neural Networks (GNNs) for performance prediction and resource allocation. It is structured into three main parts: improvement of an existing GNN model for Key Performance Indicator (KPI) prediction, dataset creation for network slicing, and the creation of a GNN model for predicting network slicing KPIs. The ultimate goal of this work is to build a model for predicting network slicing KPIs. GNNs models are a novel and powerful technique for accurately learning from graph-structured data, making them suitable for predicting network KPIs. To learn GNNs programming, the first part of this work describes the participation in a ITU challenge. Autonomous network management is explored, being essential for the dynamic environments expected in B5G networks. The limitations of traditional modeling tools and network simulators are also explored, proposing GNNs as an effective alternative due to their high accuracy and low computational requirements. A significant contribution is the enhancement of the RouteNet baseline model, achieving an improvement in prediction accuracy for larger networks, in comparison to the networks seen during training. As the goal is to build a GNNs model for predicting network slicing KPIs, and a lack of data containing network slicing scenarios is identified, the second part presents a the creation of a network slicing dataset designed to support Artificial Intelligence (AI)-based performance prediction in B5G networks. This dataset, generated through a packet-level simulator, includes diverse network scenarios with varying topologies, slice instances, and traffic flows, capturing the complexities of Enhanced Mobile Broadband (eMBB), Ultra-Reliable Low-Latency Communication (URLLC), and Massive Internet of Things (mIoT) slices. The dataset is a valuable resource for the research community, facilitating innovations in network slicing and resource management. After creating the required data, the GNN model called GNNetSlice is developed in part three, introducing a novel model that leverages GNNs to predict the performance of network slices in the core and transport network. By adopting a data-driven approach, GNNetSlice balances prediction speed and accuracy. The model demonstrates high accuracy in predicting delay, jitter, and losses across various scenarios. Overall, this thesis makes contributions to the field of network slicing, providing tools and datasets for efficient and accurate KPI prediction in B5G networks. The proposed models and datasets pave the way for more resilient and adaptive network management solutions, crucial for the next generation of mobile networks