Leveraging edge-aware graph neural networks to predict node load in backbone networks

Abstract Graph Neural Networks (GNNs) are effective machine learning (ML) techniques to study network data, particularly for large and complex systems. Many state-of-the-art models mainly consider properties of nodes and do not take into account that the structural and functional properties of edges can efficiently affect the behavior of a system. In this work, we develop and evaluate AttEAGNN, an attention-based, edge-aware GNN architecture designed for incorporating edge information when predicting network dynamics. We apply our model to the practical problem of predicting node load in two real-world backbone networks. By creating a hybrid architecture that processes nodes and edges in parallel and encodes both explicit link attributes and implicit topological features (e.g. edge centrality), our model generates a more robust representation of link-based structure. Our results demonstrate that this edge-aware approach consistently achieves higher performance on this task when compared to state-of-the-art GNNs that do not integrate rich edge features.