Graph-Based Learning for Intelligent Network Management: Adaptive Restoration, Reliability Prediction, and Traffic Forecasting

Abstract

Modern communication networks are becoming increasingly complex due to rising traffic demands, stringent reliability requirements, and the need for adaptive management across optical and wireless infrastructures. Traditional rule-based and static management approaches are often inadequate for handling the structural, temporal, and operational complexity of these environments. This thesis investigates Graph Neural Networks (GNNs) as a unifying learning paradigm for intelligent network management, and demonstrates how task-specific integration with reinforcement learning, temporal sequence models, and evolutionary optimization can improve decision-making across diverse networking problems. Three representative network management challenges are addressed. First, a GNN-enhanced deep reinforcement learning framework is developed for dynamic restoration in Wavelength Division Multiplexing (WDM) networks. By encoding topology and failure states as graph-structured inputs, the framework enables topology-aware routing and wavelength assignment, achieving faster and more effective recovery than conventional heuristic restoration methods. Second, a graph-based predictive framework is proposed for proactive radio link failure prediction in 5G networks. By combining spatial relationships among neighboring sites with temporal and environmental context, the model improves the early detection of service disruptions and supports proactive reliability management. Third, a spatio-temporal GNN-based traffic forecasting model, augmented with Transformer learning and Genetic Algorithm-based hyperparameter optimization, is introduced for 6G network environments. This model captures both structural and temporal traffic dependencies, improving forecasting accuracy and enabling more effective predictive resource allocation. Taken together, these contributions show that graph-based learning provides a coherent and effective foundation for intelligent network management across resilience, reliability, and forecasting tasks. Rather than proposing a single universal architecture for all network problems, this thesis demonstrates how a common graph-centered perspective can be adapted to different management objectives through appropriate learning and optimization mechanisms. The results establish GNN-based modeling as a practical pathway toward more adaptive, predictive, and autonomous communication networks.