Null importance: Disentangling relevance for interpretable machine learning
Abstract
Feature importance is central to interpretable machine learning, but the term "importance" encompasses several fundamentally different notions of relevance. We develop a unified perspective based on null importance: a population-level characterization of when a feature is irrelevant under a specified notion of relevance. We consider standard notions of null importance arising from marginal and conditional statistical relevance, predictive risk, functional invariance, and causal effects, and show how these notions answer different scientific questions. We illustrate the framework in two applications in which the distinction is particularly consequential: algorithmic fairness, where common fairness criteria correspond to different notions of null importance, and genomic perturbation modeling, where different notions of relevance lead to different conclusions about what a prediction model has learned. The framework connects three aspects of feature analysis: the scientific question defining relevance, the data and model assumptions that shape how different null notions relate, and the methods used to assess importance. We establish sufficient conditions under which null notions coincide and give counterexamples showing how they diverge when those conditions fail. We then characterize which nulls different method families target and when their zero-importance statistics identify those targets. Finally, simulations spanning feature dependence, redundancy, nonlinearity, hidden features and other standard phenomena, along with case studies on image and multiomics data, provide empirical evidence for these theoretical distinctions and their practical consequences. Taken together, these results provide a common statistical language for relating scientific questions, data-generating assumptions, and algorithms, and clarify the conclusions that feature-importance analyses can support.
Related
- Mastering Model Fit: Applications of the InterModel Vigorish for Interpretable Machine Learning
- Multiscale Nonlinear Risk Transmission in the Carbon–Fossil Energy–Clean Energy System from a Fractal-Market Perspective: Evidence from EMD–TVP-VAR and Interpretable Machine Learning
- An Interpretable Decision Modeling and Machine Learning-Driven Method for Bio-Inspired Styling Generation of New Energy Vehicles
- Predicting mental health risk in adolescents using interpretable machine learning: Evidence from SHAP analysis
- Development and external evaluation of an interpretable machine-learning model for early prediction of organ failure in higher-risk acute pancreatitis patients: A multicentre cohort study
BibTeX
@misc{raskutti2026null,
title = {{Null importance: Disentangling relevance for interpretable machine learning}},
author = {Garvesh Raskutti and Kris Sankaran and Jiaxin Ye},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.19511},
eprint = {2609.19511},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.19511},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Null importance: Disentangling relevance for interpretable machine learning
AU - Garvesh Raskutti
AU - Kris Sankaran
AU - Jiaxin Ye
PY - 2026
DA - 2026/09/16
PB - arXiv
DO - 10.48550/arxiv.2609.19511
UR - https://arxiv.org/abs/2609.19511
AB - Feature importance is central to interpretable machine learning, but the term "importance" encompasses several fundamentally different notions of relevance. We develop a unified perspective based on null importance: a population-level characterization of when a feature is irrelevant under a specified notion of relevance. We consider standard notions of null importance arising from marginal and conditional statistical relevance, predictive risk, functional invariance, and causal effects, and show how these notions answer different scientific questions. We illustrate the framework in two applications in which the distinction is particularly consequential: algorithmic fairness, where common fairness criteria correspond to different notions of null importance, and genomic perturbation modeling, where different notions of relevance lead to different conclusions about what a prediction model has learned. The framework connects three aspects of feature analysis: the scientific question defining relevance, the data and model assumptions that shape how different null notions relate, and the methods used to assess importance. We establish sufficient conditions under which null notions coincide and give counterexamples showing how they diverge when those conditions fail. We then characterize which nulls different method families target and when their zero-importance statistics identify those targets. Finally, simulations spanning feature dependence, redundancy, nonlinearity, hidden features and other standard phenomena, along with case studies on image and multiomics data, provide empirical evidence for these theoretical distinctions and their practical consequences. Taken together, these results provide a common statistical language for relating scientific questions, data-generating assumptions, and algorithms, and clarify the conclusions that feature-importance analyses can support.
ER -