The paper introduces a new visualization method called Accumulated Local Effects (ALE) plots to address the limitations of Partial Dependence (PD) plots in black box supervised learning models. PD plots, while popular, can produce erroneous results when predictors are strongly correlated due to their reliance on extrapolation beyond the training data. ALE plots, on the other hand, avoid this issue by using conditional densities instead of marginal densities, making them less computationally expensive and more reliable. The authors define ALE main and second-order effects for individual predictors and pairs of predictors, and provide estimators for these effects that are consistent and computationally efficient. They demonstrate the effectiveness of ALE plots through toy examples and a real-world data set, showing that ALE plots provide more accurate and interpretable results compared to PD plots, especially in the presence of highly correlated predictors. The paper also discusses the computational advantages of ALE plots over PD plots and their ability to capture complex interactions between predictors.The paper introduces a new visualization method called Accumulated Local Effects (ALE) plots to address the limitations of Partial Dependence (PD) plots in black box supervised learning models. PD plots, while popular, can produce erroneous results when predictors are strongly correlated due to their reliance on extrapolation beyond the training data. ALE plots, on the other hand, avoid this issue by using conditional densities instead of marginal densities, making them less computationally expensive and more reliable. The authors define ALE main and second-order effects for individual predictors and pairs of predictors, and provide estimators for these effects that are consistent and computationally efficient. They demonstrate the effectiveness of ALE plots through toy examples and a real-world data set, showing that ALE plots provide more accurate and interpretable results compared to PD plots, especially in the presence of highly correlated predictors. The paper also discusses the computational advantages of ALE plots over PD plots and their ability to capture complex interactions between predictors.