AttEAGNN: Attention Based Edge-Aware GNN Applied to Network Load Prediction

• Wagner Almeida UFV
• Fábio Ramos UFV
• Ricardo dos Santos UFV
• José Augusto Nacif UFV
• Alex Borges UFJF

Abstract

Graph Neural Networks (GNNs) are deep machine learning methods designed to solve tasks on complex datasets structured as graphs. Most GNNs, however, usually focus on node or whole graph representation, overlooking edge features and edge structural relationships. This paper presents AttEAGNN, an attention-based edge-aware GNN model to predict node loads in a backbone network. The proposed model can process implicit and explicit edge and node features, contributing to improved data representation. Aiming to optimize a real-world backbone network, we apply AttEAGNN to analyze and predict node loads. Our model outperforms the results of network node load predictions obtained by different state-of-the-art baselines by up to 3%. The framework we developed is publicly available for reproduction and evaluation.

Keywords: machine learning, graph neural network, edge-aware graph neural network, network optimization