Topology-Aware vs Feature-Based QoT Estimation in Multiband Optical Networks

Accurate estimation of the Signal-to-Noise Ratio (SNR) of prospective lightpaths is a key requirement for the efficient operation of dynamic multiband optical networks. This paper compares a topology-aware approach based on Graph Neural Networks (GNNs) to a traditional feature-based eXtreme Gradient Boosting (XGBoost) regressor for SNR prediction. The GNN exploits knowledge of the complete network topology and per-link spectral occupation, while XGBoost relies on a compact set of features describing the closest spectral environment along the selected path. The two approaches are evaluated under an incremental traffic-loading scenario, in which the C band is filled firstly, and the L band is later activated when needed. Results show that the GNN provides higher accuracy and more balanced prediction errors during early network operation, when spectral occupation is sparse and limited training data are available. As the network evolves toward full multiband utilization, XGBoost achieves lower mean absolute and mean squared errors, but exhibits systematic overestimation of SNR for long-haul lightpaths. In contrast, the GNN maintains better-calibrated predictions, especially for long lightpaths, as reflected by lower mean and median errors.