Impossibility Theorems for Feature Attribution

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Bilodeau, Blair, Jaques, Natasha, Koh, Pang Wei, Kim, Been
Natura: Preprint
Pubblicazione: 2022
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866916082453839872
author Bilodeau, Blair
Jaques, Natasha
Koh, Pang Wei
Kim, Been
author_facet Bilodeau, Blair
Jaques, Natasha
Koh, Pang Wei
Kim, Been
contents Despite a sea of interpretability methods that can produce plausible explanations, the field has also empirically seen many failure cases of such methods. In light of these results, it remains unclear for practitioners how to use these methods and choose between them in a principled way. In this paper, we show that for moderately rich model classes (easily satisfied by neural networks), any feature attribution method that is complete and linear -- for example, Integrated Gradients and SHAP -- can provably fail to improve on random guessing for inferring model behaviour. Our results apply to common end-tasks such as characterizing local model behaviour, identifying spurious features, and algorithmic recourse. One takeaway from our work is the importance of concretely defining end-tasks: once such an end-task is defined, a simple and direct approach of repeated model evaluations can outperform many other complex feature attribution methods.
format Preprint
id arxiv_https___arxiv_org_abs_2212_11870
institution arXiv
publishDate 2022
record_format arxiv
spellingShingle Impossibility Theorems for Feature Attribution
Bilodeau, Blair
Jaques, Natasha
Koh, Pang Wei
Kim, Been
Machine Learning
Artificial Intelligence
Despite a sea of interpretability methods that can produce plausible explanations, the field has also empirically seen many failure cases of such methods. In light of these results, it remains unclear for practitioners how to use these methods and choose between them in a principled way. In this paper, we show that for moderately rich model classes (easily satisfied by neural networks), any feature attribution method that is complete and linear -- for example, Integrated Gradients and SHAP -- can provably fail to improve on random guessing for inferring model behaviour. Our results apply to common end-tasks such as characterizing local model behaviour, identifying spurious features, and algorithmic recourse. One takeaway from our work is the importance of concretely defining end-tasks: once such an end-task is defined, a simple and direct approach of repeated model evaluations can outperform many other complex feature attribution methods.
title Impossibility Theorems for Feature Attribution
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2212.11870