Neural Graphs: an Effective Solution for the Resource Allocation in NFV Sites interconnected by Elastic Optical Networks

Modern communication networks are no longer static collections of routers and servers. They are dynamic ecosystems where virtual network functions are continuously instantiated, migrated, and resized according to changing traffic demands. In these environments, deciding how many computational resources should be allocated to each virtual function becomes a complex prediction problem that directly affects both service quality and operational efficiency. This work explores a different way of approaching this challenge. Instead of treating each virtual network function as an isolated time series, the proposed methodology models the entire infrastructure as a graph, allowing artificial intelligence to learn not only how resource demand evolves over time, but also how different network functions influence one another through their connectivity.

The proposed framework operates within a Network Function Virtualization (NFV) architecture interconnected by Elastic Optical Networks. Monitoring agents periodically collect the processing capacity required by every Virtual Network Function Instance (VNFI), while a centralized orchestrator gathers these measurements and constructs a graph representing the current state of the infrastructure. In this graph, each node corresponds to a VNFI and is labelled with its measured processing demand, while the edges capture the relationships between interconnected virtual functions. Rather than analysing individual resource traces independently, the model processes the complete graph, enabling prediction to exploit both temporal evolution and network topology simultaneously. This graph-based representation allows the orchestration system to anticipate future processing requirements before resource reconfiguration decisions are taken.

At the core of the methodology lies a Spatio-Temporal Graph Convolutional Network (STGCN). Unlike conventional recurrent neural networks, which primarily learn temporal dependencies, the proposed architecture combines temporal gated convolutions with graph convolutions that explicitly model interactions among neighbouring nodes. Each prediction block performs two complementary operations. Temporal convolutions analyse how the processing demand of each VNFI changes over consecutive monitoring intervals, while Chebyshev graph convolutions propagate information across neighbouring nodes in the graph, capturing spatial dependencies created by the virtualized network topology. These spatio-temporal convolution blocks are stacked together and followed by a final linear prediction layer, creating a compact model capable of forecasting the next processing capacity for every node simultaneously.

The methodology is evaluated using a realistic NFV scenario based on the USA backbone network, where multiple NFVI Points of Presence are interconnected through an Elastic Optical Network. The resulting virtual infrastructure generates a graph containing 48 VNFI organised into 5 connected subgraphs. Processing capacities are monitored every ten minutes over four weeks, creating time series used to train both the proposed graph neural network and a conventional LSTM-based predictor. The comparison focuses on two key performance indicators: prediction accuracy and training complexity. The results show that the graph-based approach achieved prediction errors comparable to, or better than, the LSTM model across the different subgraphs, while dramatically reducing training time. In every evaluated scenario, the STGCN requires only about 24 seconds for training, whereas the LSTM architecture required between approximately 136 and 266 seconds.

The significance of this work extends beyond faster forecasting. It introduces a different perspective on how artificial intelligence can manage communication infrastructures. Rather than analysing independent resource streams, the network itself becomes the learning structure, allowing relationships among virtualized functions to contribute directly to the prediction process. This ability is particularly relevant for future cloud-native and optical networking environments, where resource allocation, service orchestration, and network topology are deeply interconnected. By combining graph representations with temporal learning, the proposed solution demonstrates that infrastructure-aware AI can deliver accurate resource prediction while significantly reducing computational complexity. As virtualized networks continue to increase in size and complexity, graph neural networks may become one of the key technologies enabling scalable, intelligent, and proactive orchestration of next-generation communication systems.