Improving interpretability in machine learning using confidence intervals in ALE plots

Machine learning models that predict a response based on a collection of predictor variables usually do not provide simple numeric summaries of predictor effects and so are often referred to as black box models. Accumulated local effects (ALE) plots have been developed to allow visual interpretability of predictor effects in such black box models. We present R package AleCI, which improves the original ALE implementation by adding a bootstrapped confidence interval around each prediction showing the range where the true value of the predictor’s effect should exist, for categorical as well as continuous predictors. AleCI is applicable across a variety of machine learning models and updates the graphing capabilities of the original implementation by using ggplot2.