Forecasting Auxiliary Energy Consumption for Electric Heavy-Duty Vehicles

One-size-fits-all AI models can be misleading: The paper demonstrates that global explanation methods like LIME and SHAP can produce misleading results when applied to heterogeneous populations (e.g., different vehicle types, drivers, or conditions), a phenomenon related to Simpson's Paradox where overall trends contradict trends within subgroups.

Segmentation improves both accuracy and trustworthiness: Training separate models for different sub-populations improved both prediction performance and the consistency of explanations, suggesting that breaking complex problems into simpler, population-specific models enhances both technical performance and user trust.

Explainability matters for real-world deployment: The paper emphasizes that "understanding why certain predictions are cast is paramount for such a predictive model to gain user trust and be deployed in practice," highlighting that accurate predictions alone are insufficient—stakeholders need to understand the reasoning behind decisions.

Context-specific explanations are more relevant: When models are tailored to specific sub-populations (e.g., different vehicle types or operating conditions), the explanations become more consistent and meaningful within each group, making the AI system more interpretable to domain experts and operators.

Responsible AI requires acknowledging population differences: Rather than treating all users or operating conditions equally, responsible AI systems should recognize that dependencies between inputs and outcomes differ across sub-populations and adjust their models and explanations accordingly.