A GNN-Based Framework for Multi-Objective Telecommunications Quality Ranking

Assessing telecommunications mobile services quality across municipalities with heterogeneous structural, geographic, and socioeconomic conditions requires methodologies beyond simple scalar or functional aggregations. This paper introduces a self-supervised Graph Neural Network (GNN) based ranking framework that learns the geometry of municipality-level Quality of Service (QoS) indicators on a \(k\)-nearest-neighbor graph, producing a unified ranking obtained by self-supervised learning on Pareto-inspired proxies. The architecture employs residual graph convolutional layers and is trained with a multi-objective loss function that enforces pairwise ranking consistency, input reconstruction fidelity, geometric smoothness on the learned manifold, and directional alignment with a Pareto-inspired proxy. The approach is validated on a dataset covering Brazilian municipalities over multiple monthly periods, totaling over 82,000 municipality--month observations. Results show that the proposed model produces scores with a wider effective range (\(\sigma=0.098\)) than ANATEL's legacy QoS (\(\sigma=0.054\)) and its revised version (\(\sigma=0.058\)) baselines, improving rank discrimination. Distribution and spatial analyses suggest that the framework produces more informative and distinguishable municipality rankings than the baselines, mitigating the concentration effects observed in current regulatory composite indicators.