Beyond model-specific biases: An explainable multifaceted approach for robust PM₁₀ source apportionment.

Area of Science:

• Environmental Science
• Data Science
• Computational Chemistry

Background:

• PM10 source apportionment is crucial for urban air quality management.
• Integrating Positive Matrix Factorization with machine learning (ML) models like XGBoost, Random Forest (RF), and Support Vector Machine (SVM) offers a novel approach.
• The Lung Performance Optimization (LPO) algorithm and 10-fold cross-validation were used to enhance model robustness.

Discussion:

• LPO-XGBoost achieved the highest predictive accuracy (r² = 0.88).
• SHAP values were used for feature importance interpretation, but model-specific biases (e.g., early feature selection in tree-based models, kernel transformations in SVM) raise concerns.
• Variability in feature importance across models may indicate methodological inconsistencies rather than analytical depth, potentially undermining insights in noisy, high-dimensional environmental data.

Key Insights:

• Machine learning models, particularly LPO-XGBoost, demonstrate strong predictive capabilities for PM10 source apportionment.
• Model-dependent biases in SHAP value interpretations complicate the assessment of feature importance.
• The reliability of source apportionment findings is challenged by inherent model biases and data complexities.

Outlook:

• Future research should incorporate unsupervised learning and non-parametric methods (e.g., Feature Agglomeration, Highly Variable Gene Selection, Spearman's rho, Kendall's tau) for improved interpretability and robustness.
• These alternative methods can better capture complex associations and reduce the influence of model-specific biases.
• A more systematic evaluation approach is needed to strengthen reproducibility and support informed environmental decision-making.