Addressing RouteNet scalability through input and output design

With recent advances in the field of Machine Learning (ML), a multitude problems related to communication systems and networks can be solved with data-driven solutions. Since data these is mostly represented as graphs, Graph-based Neural Networks (GNNs) are good candidate for solving such problems. These GNNs used computer network modeling technique build models that accurately estimate Key Performance Indicators (KPI) delay or jitter real scenarios order ensure their requirements terms service assurance. To GNN solutions higher accuracy, low computational resource requirements, easy deployment synthetic training results into real-world networks, it more than necessary develop efficient effective models. This paper presents model capable estimating average per flow networks. By designing scale-independent features using notions from queuing theory, proposed successfully generalizes large size topologies, routing configurations, traffic matrices not seen during phase.