First study to explicitly verify predictor-target independence before classification in ML-based carbon emission analysis. While prior work (Serafeim & Velez Caicedo, Wang & Ye, Safaei et al.) builds predictive models directly from emission-related datasets without testing whether features are structurally independent of the target, this study confirms — via correlation analysis, chi-square testing (χ² = 739.33, p < 0.001), and ANOVA (F = 151.90, p < 0.001) — that its two predictors (sector membership and trade margin behavior) are genuinely independent of the emission intensity label before training. This directly addresses a data leakage risk demonstrated in the paper itself: a finer NAICS code raised accuracy to 87.75% but was discarded after diagnostics showed 98.6% train-test overlap, proving memorization rather than a learnable pattern.
Secondary contributions:
1. First sector-level (not firm-level or route-level) profiling of the full EPA Supply Chain GHG Emission Factors dataset (1,016 industries, 19 sectors)
2.Six-classifier benchmark validated against dummy baselines with a McNemar test (χ² = 53.17, p < 0.001), confirming Random Forest’s 66.14% accuracy reflects genuine signal, not chance
3. Class-specific driver analysis via SHAP showing sector identity and margin behavior contribute differently by emission class (margins dominate for High, sector dominates for Low)
