Explainable TabNet-based model for vehicular CO2 emission prediction with advanced metaheuristic optimization

Accurate and interpretable prediction of vehicular carbon dioxide (CO 2 ) emissions is essential for advancing sustainable transportation and carbon-neutral mobility. This study presents a controlled experimental framework to compare the interpretability, attention allocation, feature sparsity, and learning behavior of TabNet configurations selected by Halfway Escape Optimization (HEO) and Firefighter Optimization (FFO) for vehicular CO 2 emission prediction. Rather than focusing solely on predictive accuracy, the study examines differences in feature-attribution patterns, attention concentration, and sparsity behavior between the HEO- and FFO-selected TabNet configurations. The results indicate that the TabNet configurations selected by HEO and FFO exhibit distinct attention-allocation patterns, feature-sparsity structures, and explanation behaviors. While high predictive accuracy is expected due to the strong physical relationship between fuel-consumption variables and CO 2 emissions, the primary contribution of this study lies in characterizing differences in interpretability and learning behavior between the HEO- and FFO-selected TabNet configurations. Feature-ablation analysis further confirms this dependence: after removing all fuel-consumption variables, R 2 decreased from above 0.99 to approximately 0.94 for both optimized configurations, while substantial predictive information remained in the non-fuel vehicle attributes. Regression Error Characteristic (REC) analysis shows that approximately 93% of test samples fall within a ±10 g/km error threshold. Interpretability analysis indicates that fuel consumption variables and engine size dominate feature attribution, consistent with known physical relationships in the dataset. The alignment between TabNet feature masks and SHAP attributes reflects consistency in feature importance patterns across interpretability methods, rather than causal inference. Overall, this study provides insights into interpretability differences across optimizer-selected TabNet configurations.