An Approximate Solution Using Graph Convolutional Network for a Load Balancing Problem in Communication Networks

The rapid increase in data traffic from smartphones, cloud computing, and IoT devices has led to severe congestion in communication networks. In such networks, the load balancing problem—formulated as the Unsplittable flow Edge Load factor Balancing (UELB) problem—is NP-hard, necessitating efficient approximate solution methods. This paper presents a heuristic approach that employs a Graph Convolutional Network (GCN) to predict flow distributions, combined with a beam search strategy for rapid determination of feasible routing paths. Extensive experiments on Newman–Watts–Strogatz small-world graphs demonstrate reductions in computation time compared to reinforcement learning-based K-shortest paths (RL-KSP), while maintaining competitive approximation rates. Future research will explore further improvements in accuracy and scalability.