AUTO-CONFIGURED EXPLAINABLE GRAPH NEURAL NETWORKS FOR MULTI-SITE POLLUTION PREDICTION

Research Approach

This paper addresses the limitations of static, predefined graph structures in traditional Graph Neural Networks (GNNs) when applied to air pollution forecasting. The authors propose a novel framework that dynamically constructs graph adjacency matrices using a confusion matrix derived from a supervised learning process. This allows the model to adaptively capture inter-site relationships rather than relying on fixed geographic distances. To stabilize training and mitigate the vanishing gradient problem, the authors implement a hybrid loss function that combines energy distance and Huber loss. The framework is benchmarked against five GNN architectures (GCN, SGConv, GIN, GAT, and GraphSage) and compared against traditional machine learning and deep learning baselines (Prophet, LSTM, GRU).

Key Findings

The experimental results demonstrate that the proposed dynamic graph construction and hybrid loss function significantly enhance predictive accuracy for particulate matter (PM1, PM10, and PM2.5). Among the tested GNN models, GraphSage consistently achieved the highest accuracy across various time horizons. The study also integrates GNNExplainer and PGExplainer to provide transparency into the model's decision-making process, allowing researchers to interpret how specific features and graph structures influence pollution predictions.

Accurate, localized air quality forecasting is essential for public health interventions and environmental policy. By moving away from static graph assumptions, this research provides a more flexible and robust methodology for modeling complex, non-linear spatiotemporal dependencies in environmental data. The inclusion of explainability tools further bridges the gap between high-performance black-box models and the need for actionable, transparent insights in urban planning and health risk management.