Attribution-based Explanations that Provide Recourse Cannot be Robust
Hidde Fokkema, Rianne de Heide and Tim van Erven
Attribution-based Explanations that Provide Recourse Cannot be Robust
Hidde Fokkema, Rianne de Heide and Tim van Erven
Journal of Machine Learning Research, 24(360), pp.1-37, 2023. arxiv proc
What is this paper about?
This paper studies explanation methods for machine-learning decisions from the viewpoint of actionable recourse. Its central result is an impossibility trade-off: a single local attribution method cannot, in general, be both robust to small input changes and sufficiently sensitive to recourse.
Summary
The paper formalizes recourse sensitivity for local feature-attribution explanations and shows an impossibility result: in general, a single attribution method cannot be both robust to small changes in the input and sensitive enough to identify useful recourse. The result covers popular explanation methods and counterfactual explanations, and the paper discusses possible ways around the impossibility, such as returning sets of explanations.