Model-Agnostic Interpretability: Effect Size Measures from Accumulated Local Effects (ALE)

Effectively interpreting relationships between variables and outcomes in statistical and machine learning (ML) models is crucial for advancing data science methodologies. This paper introduces novel metrics for accumulated local effects (ALE), designed to quantify the effects of variables in a nuanced and interpretable manner. Specifically, we propose ALE deviation (ALED) and ALE range (ALER), along with their normalized counterparts (NALED and NALER). Using a real dataset of bicycle sharing ridership, we compare these metrics through ordinary least squares (OLS) regression and a generalized additive model (GAM). Our findings highlight the advantages of these ALE-based measures in providing clear insights into variable impacts and enhancing model interpretability by providing a valuable numeric summary complement to the nuanced visualizations of ALE.