Explainable GNNs Improve Multi-Site Air Pollution Prediction.

This research presents an innovative approach to predicting particulate matter (PM) concentrations across multiple sites, a critical task for air pollution mitigation. The core of the method involves an auto-configured and explainable Graph Neural Network (GNN) framework. Unlike traditional GNNs that rely on predefined graphs, this system dynamically constructs its graph based on a confusion matrix derived from a supervised learning process, allowing it to better capture complex inter-class relationships in the data. To further enhance learning stability and address issues like vanishing gradients, the study incorporates a hybrid loss function combining energy distance and Huber loss. The framework was rigorously evaluated using real-world air pollution data from Salt Lake City, Utah, comparing five different GNN models. Results consistently showed that GraphSage achieved the highest prediction accuracy for PM1, PM10, and PM2.5 concentrations across various time horizons, outperforming both traditional machine learning and deep learning models. Furthermore, the application of GNNExplainer and PGExplainer provided crucial insights into feature importance and graph structure, ensuring the model's transparency and trustworthiness.