| English | 中文 |
The source is the final published article Interpretable machine learning for ambient air pollution modeling: A systemic review, Atmospheric Research 336 (2026), 108869. DOI: 10.1016/j.atmosres.2026.108869.
The methodological mapping was reviewed on 2026-10-27 against the final published version.
| Published material | Reusable implementation | Boundary |
|---|---|---|
| Section 4.2, feature engineering | Training-only missingness, variation, correlation screening and VIF diagnostics | Not an automatic causal adjustment selector |
| Section 4.3 and Figure 2 | Explicit treatment/outcome/adjustment roles, DAG checks and cross-fitted PLR | Identification assumptions remain declared, not empirically verified |
| Section 4.4 and Figures 3–6 | Performance importance, ICE/PDP and empirical-background permutation Shapley estimates | Predictive explanation; H-statistics and the article’s full figure reproduction remain separate adaptations |
| Sections 5.1–5.2 | Explicit interpretation scope, correlated-feature limitations and scientific review | A lower prediction error does not establish a causal mechanism |
Code is independently written. The continuous-treatment PLR score and sample-splitting design additionally follow the primary DoubleML model and resampling documentation. This implementation is not the DoubleML package and does not claim numerical equivalence to the paper’s estimates.
Synthetic demonstrations have a known data-generating effect. Observational AQ-Bench causal analysis requires a study-specific graph, variables, spatial dependence strategy, units and scientific assumptions; the predictive example does not silently activate that analysis.