Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models

TL;DR

This paper introduces accumulated local effects (ALE) plots, a new visualization method for predictor effects in black box models that avoids the pitfalls of partial dependence and marginal plots, providing reliable and efficient insights.

Contribution

The paper proposes ALE plots, a novel visualization technique that overcomes the limitations of PD and M plots in dependent predictor settings, with improved reliability and computational efficiency.

Findings

ALE plots do not require extrapolation like PD plots.

ALE plots are less biased than M plots when predictors are dependent.

ALE plots are computationally more efficient than PD plots.

Abstract

When fitting black box supervised learning models (e.g., complex trees, neural networks, boosted trees, random forests, nearest neighbors, local kernel-weighted methods, etc.), visualizing the main effects of the individual predictor variables and their low-order interaction effects is often important, and partial dependence (PD) plots are the most popular approach for accomplishing this. However, PD plots involve a serious pitfall if the predictor variables are far from independent, which is quite common with large observational data sets. Namely, PD plots require extrapolation of the response at predictor values that are far outside the multivariate envelope of the training data, which can render the PD plots unreliable. Although marginal plots (M plots) do not require such extrapolation, they produce substantially biased and misleading results when the predictors are dependent, analogous to the omitted variable bias in regression. We present a new visualization approach that we term accumulated local effects (ALE) plots, which inherits the desirable characteristics of PD and M plots, without inheriting their preceding shortcomings. Like M plots, ALE plots do not require extrapolation; and like PD plots, they are not biased by the omitted variable phenomenon. Moreover, ALE plots are far less computationally expensive than PD plots.