Modeling and analysis of graph neural networks for optimizing routing in infocommunication networks

The research object is the process of route construction in infocommunication networks. A routing model based on a graph neural network (GNN) with edge-level classification has been developed, implemented, and experimentally studied. The model uses the GENConv architecture. Architectural features of GNNs are analyzed, including message passing, attention mechanisms, feature aggregation, and updating techniques. A comparison of GCN, GAT, and GENConv architectures is conducted, and GENConv is justified as the base architecture for edge-level classification of routing edges. A model is built using GENConv with an MLP decoder. The model is trained on a large set of graphs and evaluated in terms of accuracy, average latency, and the percentage of successfully constructed routes. The proposed approach is compared with classical algorithms based on decision quality and execution time. It is established that in inference mode, the graph model works significantly faster – especially on large graphs – and does not require exhaustive route space reprocessing for each query. This makes the proposed method suitable for real-time use in dynamic networks, where decision-making speed is critical.