Familywise Feature Importance Stability in Chemical and Materials Machine Learning

Abstract

While feature importance analysis in chemical and materials machine learning can often be sensitive to both the predictive model and the attribution rule, the robustness of these rankings is rarely quantified before they are used to gain physical insights. Here, we compare 26 feature importance pipelines spanning data-driven, model-based, and formula-based analyses on a metal-support interaction data set anchored by an explicit SISSO equation, and we examine whether the same qualitative behavior recurs in high-entropy-alloy and halide perovskite data sets. Across the three benchmarks, we observe high intrafamily agreement but substantial interfamily variance. While a small subset of features remains stable across multiple families, several midranked features are highly family dependent, with their apparent importance shifting according to the underlying modeling assumptions. To ensure robust interpretability, we recommend that feature importance be reported by method family or correlation-based clusters, supplemented by resampling intervals.