Total Effects with Constrained Features - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2023

Total Effects with Constrained Features

Résumé

Recent studies have emphasized the connection between machine learning feature importance measures and total order sensitivity indices (total effects, henceforth). Feature correlations and the need to avoid unrestricted permutations make the estimation of these indices challenging. Additionally, there is no established theory or approach for non-Cartesian domains. We propose four alternative strategies for computing total effects that account for both dependent and constrained features. Our first approach involves a generalized winding stairs design combined with the Knothe-Rosenblatt transformation. Our second approach is a U-statistic estimator that combines the Jansen intuition with a weighting factor. The U-statistic framework allows the derivation of a central limit theorem for this estimator. However, this design is computationally intensive. Then, our third approach uses derangements to significantly reduce computational burden. We prove consistency and central limit theorems for these estimators as well. Our fourth approach is based on a nearest-neighbour intuition and it further reduces computational burden. We test these estimators through a series of increasingly complex computational experiments with features constrained on compact and connected domains (circle, simplex), non-compact and non-connected domains (Sierpinski gaskets), and conclude with an application to a realistic simulator.
Fichier principal
Vignette du fichier
Totals.pdf (2.89 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04102882 , version 1 (22-05-2023)
hal-04102882 , version 2 (06-02-2024)

Identifiants

  • HAL Id : hal-04102882 , version 1

Citer

Emanuele Borgonovo, Elmar Plischke, Clémentine Prieur. Total Effects with Constrained Features. 2023. ⟨hal-04102882v1⟩
92 Consultations
72 Téléchargements

Partager

Gmail Facebook X LinkedIn More