This metric evaluates how robust an explanation method is to changes in its configuration. Since many XAI methods rely on hyperparameters, high sensitivity may indicate instability, making tuning more difficult and reducing user trust in the method [Bansal et al. (2020)].
A common approach is to generate explanantia for the same input across different hyperparameter settings and measure the similarity between them [Verma and Ganguly (2019), Bansal et al. (2020), Mishra et al. (2020), Sanchez-Lengeling et al. (2020), Graziani et al. (2021)]. Alternatively, [Chen et al. (2019b)] propose to compare the stability of performance metrics such as fidelity over varying hyperparameters. Beyond identifying general sensitivity, [Graziani et al. (2021)] use this technique to guide hyperparameter selection: by progressively adjusting hyperparameters and tracking when the generated explanantia converge, one can identify stable regions in the hyperparameter space.
Although most work focuses on FAs, the general idea is applicable to any method with tunable hyperparameters.