Accumulated Local Effects (ALE) plots

ALE plots are like the PDPs we studied in Chapter 4, Fundamentals of Feature Importance and Impact, except they are unbiased and much faster. By unbiased, we mean they don't have an assumption that seldomly holds true: features are uncorrelated. As we've noticed already, co2 and ghgScore have been derived from co2TailPipeGpm. Therefore, they are mostly redundant – except when they are -1 ("Not Available"). So, how can we rely on an interpretation method that confounds their effects?

Thanks to its properties, there's a lot of consistency in SHAP's attributions because it makes simulations based on reasonable expectations – even accounting, for the most part, for colinear features. PDPs make averages of predictions across all feature values (and interpolations) regardless of whether they make sense while assuming independence of the features.

On the other hand, ALE plots take a reasonable approach...