Delay estimation based on multiple stage message passing with attention mechanism using a real network communication dataset

Modeling network communication environments with Graph Neural Networks (GNNs) has gained notoriety in recent years due to the capability of GNNs to generalize well for data defined over graphs. Hence, GNN models have been used to abstract complex relationships from network environments, creating the so-called digital twins, with the objective of predicting important quality of service metrics, such as delay, jitter, link utilization, and so on. However, most previous work has used synthetic data obtained with simulations. The research question posed by the "ITU Graph Neural Networking Challenge 2023" is whether GNN models are capable of estimating the mean per-flow delay network, using data from a real network environment. The solution presented in this paper achieved first place in the mentioned challenge. It adopted a GNN based on multiple-stage message passing and the attention mechanism to predict the mean per-flow delay. Furthermore, feature selection was used to choose a reasonable subset of input parameters. The developed GNN model achieved a mean absolute percentage error under 20.1% in the challenge test dataset, which was composed by network conditions not used in the training dataset.