Explaining Meta-learner’s Predictions: Case of Corporate CO2 Emissions

Production activities of companies very often leads to release of carbon dioxide (CO \(_2\) ) into the atmosphere which pollutes the environment. It is therefore necessary for companies to control and reduce their pollution levels, and this requires knowing the amount of CO \(_2\) emissions that can be produced and identifying the factors responsible for it. Several works have been carried out with the aim of predicting the quantity of CO \(_2\) emitted at the company level. One of these works use a Stacked Generalization ensemble model for prediction, and get better performance. This aforementioned work proposes an analysis of the role of variables in the prediction, but these explanations are provided based on some weak learners while the best prediction performance is obtained by the whole ensemble model. In this work we propose to explain the prediction of CO \(_2\) emissions from the whole Stacked Generalization ensemble model itself. For this we propose an explainability method that combines the strenghts of two state of the art explainability techniques (LRP and SHAP). Experiments show that the proposed method can be relevant both in term of stability and fidelity.