Tag
The paper proposes a new evaluation method for machine learning explanations by converting explanations into predictors and testing their ability to reproduce model predictions. It demonstrates that the effectiveness of explanation methods like SHAP and PDP depends on the independence of features in the data.