Most machine learning models are designed to be "forward-looking": they take data about a person and predict an outcome, such as the risk of heart disease or the likelihood of a loan default. However, for a person receiving a negative prediction, the most important question is not "Why did this happen?" but "What can I actually do to change it?"
This shift from explanation to intervention is the core of Algorithmic Recourse. While traditional Explainable AI (XAI) focuses on feature importance, recourse focuses on providing a feasible path for a user to reverse an unfavorable decision.
The Actionability Gap
The "original sin" of early counterfactual explanation methods was their reliance on mathematical proximity over real-world feasibility. As Timo Freiesleben and Mark T. Keane argue, a model might suggest that a loan applicant would be approved if they were five years younger or lived in a different zip code. While these are mathematically valid "counterfactuals"—points across the decision boundary—they are physically impossible for the user to implement.
To bridge this gap, researchers like Michael T. Lash and W. Nick Street proposed partitioning features into three categories:
- Immutable: Traits that cannot change (e.g., genetics, birth year).
- Directly Modifiable: Actions a user can take (e.g., increasing savings, taking a course).
- Indirectly Affected: Features that change as a result of other actions (e.g., credit score).
By enforcing strict constraints on immutable features, frameworks like CERTIFAI and MACE ensure that the advice provided to users is grounded in reality.
From Feature Tweaks to Causal Interventions
Even when features are mutable, changing them in isolation often ignores the laws of cause and effect. Amir-Hossein Karimi, Isabel Valera, and Bernhard Schölkopf argue that recourse should be framed as a causal intervention rather than a simple feature "tweak."
In their view, a user doesn't just "change" their credit score; they perform an action (like paying off a debt) that causes their credit score to change. If a model suggests increasing income without accounting for the fact that income is causally linked to education level and age, the advice may be misleading. Using Structural Causal Models (SCMs), researchers can now generate recourse that accounts for these downstream effects, ensuring that the suggested path is not just a point in space, but a viable sequence of actions.
Training for Recourse
Traditionally, recourse is treated as a post-hoc problem: we take a trained model and try to find a way around its boundaries. However, Alexis Ross and Himabindu Lakkaraju suggest that we should train models to be "recourse-friendly" from the start.
If a model achieves high accuracy by relying heavily on immutable features (like age), it effectively "locks" certain users into negative outcomes. By incorporating actionability constraints into the training process—a method called Counterfactual Training—models are forced to base their decisions on features that users can actually influence. This shifts the decision boundary to prioritize actionable factors, reducing the "cost" of recourse for the end-user.
Verifying the Right to Recourse
How do we know if a model is fair? Standard fairness metrics look at parity in predictions, but Berk Ustun and Tsui-Wei Weng argue for recourse verification.
Using the concept of reachable sets—the total space of possible states a person can reach given their constraints—auditors can certify whether a model allows for recourse. If a person is stuck in a "confined region" where no possible action can flip the model's decision, they are effectively precluded from access. This approach, exemplified by the FACTS framework, allows for auditing "fairness of recourse," ensuring that the effort required to change a decision is distributed equitably across different demographic groups.
The Shift to Collective and Robust Recourse
Recent work has expanded the scope of recourse beyond the individual. Elliot Creager and Richard Zemel explore Collective Action, where groups of users coordinate to "nudge" a model’s parameters, making the system more favorable for everyone without requiring impossible individual changes.
Furthermore, researchers like Ricardo Dominguez-Olmedo have highlighted the fragility of recourse: a suggestion that works today might fail tomorrow if the model is slightly retrained. This has led to the development of Robust Recourse, which seeks to find paths that remain valid even under model uncertainty or noisy implementation by the user.
By moving from "why" to "how," the field of actionable recourse ensures that AI systems do not just judge individuals, but provide them with the agency to improve their outcomes.