How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation

Abstract

Graph neural networks (GNNs) are increasingly used to estimate air pollution at unmonitored urban locations, but reported gains often conflate what a model is with what it sees. We present a heterogeneous, wind-aware graph attention network that embeds regulatory monitoring stations and traffic counters as distinct node types, and evaluate it under an information-parity protocol: six geostatistical and tabular baselines receive the identical, published information set, within a task taxonomy separating virtual sensing (target history available) from spatial extrapolation (no target-side data at any stage). On five stations and 16 traffic counters in Graz, Austria (2018-2021), neighboring stations’ same-day concentrations account for essentially the entire previously observed GNN advantage: the architecture gap at parity is -0.002 R2. In leakage-free extrapolation the GNN does not outperform task-legal geostatistics (-0.06 R2 versus the strongest baseline), a heterogeneous GNNExplainer analysis routes almost all explanation mass through station-to-station rather than traffic edges, and split-conformal intervals under-cover (0.53 and 0.76 at nominal 0.90), quantifying the cost of crossstation calibration. In dense daily-resolution urban networks, evaluation practice rather than architecture dominates reported GNN advantages; the protocol, code, and data are released so the audit can be repeated on any network.