References:
Wachter et al. (2017), Karimi et al. (2020), Mothilal et al. (2020), Nguyen and Martínez (2020), Stepin et al. (2020), Verma et al. (2024)
Toggle Text Reference
In cases where multiple counterfactuals are generated for a single instance, they should be diverse to offer distinct and meaningful alternatives [Wachter et al. (2017), Stepin et al. (2020), Verma et al. (2024)].
Diversity is typically quantified by computing pairwise distances among the set of counterfactuals, either as the average pairwise distance [Mothilal et al. (2020), Nguyen and Martínez (2020)], or as the number of counterfactual pairs that exceed a predefined distance threshold [Karimi et al. (2020)].
Diversity is typically quantified by computing pairwise distances among the set of counterfactuals, either as the average pairwise distance [Mothilal et al. (2020), Nguyen and Martínez (2020)], or as the number of counterfactual pairs that exceed a predefined distance threshold [Karimi et al. (2020)].