On the global feature importance for interpretable and trustworthy heat demand forecasting

Multiple interpretability methods provide complementary insights: The paper compares four different approaches (intrinsic Gradient Boosting interpretability, Partial Dependence, Accumulated Local Effects, and SHAP) rather than relying on a single method, demonstrating that different techniques can reveal different aspects of how ML models make predictions in heat demand forecasting.

Avoiding biased interpretation methods matters: The study specifically avoids feature permutation and perturbation approaches because these can introduce unrealistic data values that bias results, highlighting that the choice of interpretability technique itself can affect whether stakeholders can trust the explanations they receive.

Interpretability supports real-world accountability: By making heat demand forecasting models more interpretable and trustworthy, the research addresses practical concerns including regulatory compliance ("adherence to communal standards"), customer satisfaction, and liability risks—showing that explainability isn't just academic but essential for responsible deployment in critical infrastructure.

Domain context matters for interpretation: The paper emphasizes that feature importance assessments must be interpreted within the specific context of district heating processes, suggesting that responsible AI requires domain expertise and cannot rely solely on generic interpretability metrics.

Ante-hoc methods reduce explanation gaps: By using intrinsic interpretability of the model itself rather than only post-hoc explanations added afterward, the approach creates more direct connections between the model's actual decision-making process and human understanding.