The Quest for the Best Explanation: Comparing Models and XAI Methods in Air Quality Modeling Tasks
Air quality (AQ) modeling is at the forefront of estimating pollution levels in areas where the spatial representativity is low. Large metropolitan areas in Asia such as Beijing face significant pollution issues due to rapid industrialization and urbanization. AQ nowcasting, especially in dense urba...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-07-01
|
| Series: | Applied Sciences |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2076-3417/15/13/7390 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Air quality (AQ) modeling is at the forefront of estimating pollution levels in areas where the spatial representativity is low. Large metropolitan areas in Asia such as Beijing face significant pollution issues due to rapid industrialization and urbanization. AQ nowcasting, especially in dense urban centers like Beijing, is crucial for public health and safety. One of the most popular and accurate modeling methodologies relies on black-box models that fail to explain the phenomena in an interpretable way. This study investigates the performance and interpretability of Explainable AI (XAI) applied with the eXtreme Gradient Boosting (XGBoost) algorithm employing the SHapley Additive exPlanations (SHAP) and the Local Interpretable Model-Agnostic Explanations (LIME) for PM<sub>2.5</sub> nowcasting. Using a SHAP-based technique for dimensionality reduction, we identified the features responsible for 95% of the target variance, allowing us to perform an effective feature selection with minimal impact on accuracy. In addition, the findings show that SHAP and LIME supported orthogonal insights: SHAP provided a view of the model performance at a high level, identifying interaction effects that are often overlooked using gain-based metrics such as feature importance; while LIME presented an enhanced overlook by justifying its local explanation, providing low-bias estimates of the environmental data values that affect predictions. Our evaluation set included 12 monitoring stations using temporal split methods with or without lagged-feature engineering approaches. Moreover, the evaluation showed that models retained a substantial degree of predictive power (R<sup>2</sup> > 0.93) even in a reduced complexity size. The findings provide evidence for deploying interpretable and performant AQ modeling tools where policy interventions cannot solely depend on predictive analytics tools. Overall, the findings demonstrate the large potential of directly incorporating explainability methods during model development for equal and more transparent modeling processes. |
|---|---|
| ISSN: | 2076-3417 |