Abstract
The rapid densification of 5G networks and the growing diversity of service demands have made efficient and fair resource allocation increasingly challenging for network operators. Traditional optimization and heuristic game-theoretic approaches struggle to adapt to highly dynamic network conditions, often resulting in reduced throughput, increased latency, and imbalanced resource utilization. To address these issues, this paper proposes a Game-Theory-Driven Hybrid Graph Neural Network and Multi-Agent Deep Reinforcement Learning (GNN-MADRL) framework for dynamic resource allocation in multi-cell 5G networks. The framework employs GNNs to model spatial topology, interference relationships, and user-base station interactions, while MADRL agents learn adaptive strategies for bandwidth, power, and subcarrier allocation. A Nash equilibrium layer is integrated to stabilize agent interactions and ensure fairness and convergence. Experimental results on a realistic 5G QoS dataset demonstrate significant improvements over baseline methods, including higher throughput, lower latency, improved fairness, reduced packet loss, and enhanced energy efficiency, confirming the framework's scalability and QoS awareness.