This paper introduces a new visualization method, accumulated local effects (ALE) plots, for interpreting the effects of predictor variables in black box supervised learning models. Unlike partial dependence (PD) plots, which can produce unreliable results when predictors are strongly correlated, ALE plots avoid this issue by not requiring extrapolation of the response at predictor values outside the training data's range. ALE plots are also computationally more efficient than PD plots. The paper discusses the limitations of PD plots, particularly their reliance on extrapolation when predictors are correlated, leading to inaccurate visualizations. ALE plots, on the other hand, use the conditional density of the predictors to estimate the effects, avoiding the need for extrapolation. This makes ALE plots more reliable for understanding the main and interaction effects of predictors in complex models. The paper defines ALE main and second-order interaction effects for individual predictors and pairs of predictors, respectively. These effects are estimated using finite differences and summations, and are then centered to ensure they represent the true effect of the predictors without bias. The ALE decomposition of the function f(x) is shown to have an orthogonality-like property, making it suitable for visualization. The paper also provides examples demonstrating the reliability of ALE plots compared to PD plots. In a toy example with a tree model, ALE plots accurately captured the true linear and quadratic effects of the predictors, while PD plots were unreliable due to extrapolation. In a neural network example, ALE plots provided accurate estimates of the true effects, while PD plots were inaccurate. The paper further illustrates the use of ALE plots on a real-world bike-sharing dataset, showing how ALE plots can provide clear interpretations of the effects of various predictors, such as month, hour of day, weather situation, and wind speed. The ALE plots were found to be more interpretable and computationally efficient than PD plots, making them a valuable tool for understanding the effects of predictors in black box models.This paper introduces a new visualization method, accumulated local effects (ALE) plots, for interpreting the effects of predictor variables in black box supervised learning models. Unlike partial dependence (PD) plots, which can produce unreliable results when predictors are strongly correlated, ALE plots avoid this issue by not requiring extrapolation of the response at predictor values outside the training data's range. ALE plots are also computationally more efficient than PD plots. The paper discusses the limitations of PD plots, particularly their reliance on extrapolation when predictors are correlated, leading to inaccurate visualizations. ALE plots, on the other hand, use the conditional density of the predictors to estimate the effects, avoiding the need for extrapolation. This makes ALE plots more reliable for understanding the main and interaction effects of predictors in complex models. The paper defines ALE main and second-order interaction effects for individual predictors and pairs of predictors, respectively. These effects are estimated using finite differences and summations, and are then centered to ensure they represent the true effect of the predictors without bias. The ALE decomposition of the function f(x) is shown to have an orthogonality-like property, making it suitable for visualization. The paper also provides examples demonstrating the reliability of ALE plots compared to PD plots. In a toy example with a tree model, ALE plots accurately captured the true linear and quadratic effects of the predictors, while PD plots were unreliable due to extrapolation. In a neural network example, ALE plots provided accurate estimates of the true effects, while PD plots were inaccurate. The paper further illustrates the use of ALE plots on a real-world bike-sharing dataset, showing how ALE plots can provide clear interpretations of the effects of various predictors, such as month, hour of day, weather situation, and wind speed. The ALE plots were found to be more interpretable and computationally efficient than PD plots, making them a valuable tool for understanding the effects of predictors in black box models.