A Traffic-Aware Graph Neural Network for User Association in Cellular Networks

In this paper, we utilize a graph neural network to perform joint user association and access point activation/deactivation to optimize network blockage. Instead of using theoretical models to characterize the distribution of network traffic, we use a real-world network traffic dataset. In order to train the graph neural network, our method leverages reinforcement learning, specifically employing the deep deterministic policy gradient (DDPG) algorithm. This approach allows us to benefit from the advantages of both value-based and policy-based reinforcement learning methods. A simple but flexible reward function is defined to capture the trade-off between the fraction of active access points and network blockage. The graph neural network's awareness of the network topology gives the proposed method a clear performance advantage over the commonly used greedy heuristic optimization method, reducing blockage by up to more than 50% on real-world network traffic data while also reducing computational complexity from $\mathcal {O}(N^{3})$ to $\mathcal {O}(N)$. The proposed user association and access point activation method is not limited by the network architecture or technology, and can be applied to different generations of cellular networks.