How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation
8 affiliations
- 1Lisbon Council
- 2Smart Sense d.o.o
- 3Institute for Medical Research and Occupational Health
- 4Landhausgasse 7, Austria
- 5Invenium Data Insights GmbH, Herrengasse 28, Graz, Austria, 8010
- 6Institute for Anthropological Research
- 7Faculty of Food Technology Osijek
- 8Josip Juraj Strossmayer University of Osijek
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.
Related
- A comprehensive evaluation of inland water quality monitoring using the long-term Landsat archive and mixture density networks
- Blinded Evaluation of Oceanic Sound Source Locations via Sequential Bound Estimation
- How Sparse and How Noisy? Systematic Benchmarking of Inverse Physics-Informed Neural Networks for Manning Friction Estimation in Shallow Water Equations
- Cost-aware evaluation of time-series foundation models for urban air-quality and temperature forecasting
- Evaluation of HNO3, SO2, and NH3 in the Surface Tiled Aerosol and Gaseous Exchange (STAGE) option in the Community Multiscale Air Quality Model version 5.3.2 against field-scale, in situ and satellite observations
BibTeX
@misc{petric2026much,
title = {{How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation}},
author = {Valentino Petrić and Dejan Strbad and Nikolina Račić and Francesco Mureddu and Andreas Schopper and Benedikt Tschofenig and Michael Haberl and Mario Lovrić and Mario Lovric},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008206/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008206/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation
AU - Valentino Petrić
AU - Dejan Strbad
AU - Nikolina Račić
AU - Francesco Mureddu
AU - Andreas Schopper
AU - Benedikt Tschofenig
AU - Michael Haberl
AU - Mario Lovrić
AU - Mario Lovric
PY - 2026
DA - 2026/09/02
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008206/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008206/v1
AB - 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.
ER -