References:
Mahajan et al. (2019), Karimi et al. (2020), Le et al. (2020), Pawelczyk et al. (2021), Ma et al. (2022), Smyth and Keane (2022), Verma et al. (2024)
Toggle Text Reference
For a counterfactual to be helpful, it should only introduce changes that are feasible or allowed. This is often enforced by defining explicit constraints, such as
• Immutable features (e.g., sex, age) [Karimi et al. (2020), Pawelczyk et al. (2021), Smyth and Keane (2022), Verma et al. (2024)], or
• Value boundaries for individual features [Karimi et al. (2020), Le et al. (2020), Ma et al. (2022), Smyth and Keane (2022)].
The actionability of a counterfactual set can then be assessed by calculating the fraction of counterfactuals that satisfy all constraints [Karimi et al. (2020), Le et al. (2020), Pawelczyk et al. (2021), Ma et al. (2022), Smyth and Keane (2022)], or by computing the harmonic mean of satisfied constraints per counterfactual [Mahajan et al. (2019)].
• Immutable features (e.g., sex, age) [Karimi et al. (2020), Pawelczyk et al. (2021), Smyth and Keane (2022), Verma et al. (2024)], or
• Value boundaries for individual features [Karimi et al. (2020), Le et al. (2020), Ma et al. (2022), Smyth and Keane (2022)].
The actionability of a counterfactual set can then be assessed by calculating the fraction of counterfactuals that satisfy all constraints [Karimi et al. (2020), Le et al. (2020), Pawelczyk et al. (2021), Ma et al. (2022), Smyth and Keane (2022)], or by computing the harmonic mean of satisfied constraints per counterfactual [Mahajan et al. (2019)].