This paper provides a comprehensive guide to understanding and interpreting different feature importance (FI) methods for scientific inference. The authors first determine the types of feature-target associations that can be analyzed using FI methods, including unconditional association, conditional association given all remaining features, and conditional association given a user-specified set of features.
The authors then discuss three classes of FI methods: those based on univariate perturbations (permutation feature importance, conditional feature importance, and relative feature importance), those based on marginalization (marginal and conditional SAGE value functions, and SAGE values), and those based on model refitting (leave-one-covariate-out and Williamson's variable importance measure).
For each method, the authors provide interpretation guidelines based on the association types introduced earlier. They show that different FI methods provide insight into different types of associations, and that making the correct choice of FI method for a specific use case is crucial. The authors also provide mathematical results and proofs to support their interpretations.
The paper concludes by discussing options for estimating the uncertainty of FI methods and pointing to directions for future research aiming at full statistical inference from black-box machine learning models.
Egy másik nyelvre
a forrásanyagból
arxiv.org
Mélyebb kérdések