Explainability-Driven Comparison of Topology-Aware and Feature-Based QoT Estimation in Multiband Optical Networks

Accurate estimation of the Signal-to-Noise Ratio (SNR) of prospective lightpaths is a fundamental requirement for reliable operation of dynamic multiband optical networks. In recent years, a variety of machine learning approaches have been proposed for SNR prediction, ranging from topology-aware Graph Neural Networks (GNNs) to feature-based ensemble models such as eXtreme Gradient Boosting (XGB). While their predictive performance has been widely evaluated in terms of accuracy-related metrics, a systematic comparison of their decisional logic and results interpretability remain largely unexplored.This paper focuses on model explainability and offers a post-hoc interpretability study of two predictors, based on GNN and XGB, using Shapley additive explanations to examine feature importance patterns in multiband scenarios. Our analysis shows that, although XGB typically achieves lower aggregate prediction errors, it predominantly relies on few path-level descriptors. In contrast, the GNN distributes predictive influence across topology-aware embeddings and full-spectrum occupancy representations, reflecting a deeper exploitation of structural network information.