Algorithmic Recourse

Recourse must recommend changes to features that an individual can realistically control. It is insufficient to suggest a person alter an immutable characteristic or a variable outside their direct influence.

• Immutable Attributes: Age, ethnicity, or place of birth cannot be changed and must be excluded from the action set.
• Actionable Features: Recommendations should target mutable variables like income, savings balance, credit utilization, or the number of late payments.
• Constraint Awareness: Effective systems incorporate individual-specific constraints, such as a maximum feasible increase in monthly income or the inability to relocate geographically.

The recommended changes must have a genuine causal effect on the outcome. Correlational or spurious relationships can lead to 'gaming' the system without achieving the desired result.

• Structural Causal Models: Recourse algorithms should use causal graphs to distinguish between direct causes and mere correlations.
• Intervention vs. Observation: The system must predict the outcome of an intervention (doing X) rather than just observing a conditional probability.
• Avoiding Spurious Recourse: Recommending a user increase their number of open credit lines might be correlated with creditworthiness but causally harmful if it lowers the average account age.

Recourse should remain valid even if the underlying model is retrained or updated. A fragile explanation that expires after a minor model refresh is not useful for long-term planning.

• Sub-population Invariance: The causal direction of the recommendation should hold across different data distributions.
• Adversarial Robustness: The recourse path should not exploit brittle decision boundaries that disappear when the model is slightly perturbed.
• Temporal Stability: The recommended feature changes should remain valid over a reasonable time horizon, allowing the individual to execute the plan.

Effective recourse balances the number of changes required against the total 'cost' or effort for the individual. Overwhelming a user with dozens of small changes is less helpful than a few high-impact, clear steps.

• L0/L1 Norm Minimization: Algorithms often minimize the number of features changed (sparsity) to keep the plan simple.
• Cost Functions: A weighted cost matrix can encode that changing 'education level' is far more difficult than 'reducing credit card balance'.
• Diverse Counterfactuals: Providing multiple distinct paths (e.g., 'increase income by 10%' OR 'reduce debt by 20%') allows the individual to choose the most feasible route.

Recourse is inherently contrastive—it answers 'Why was I denied, and what would need to be different for an approval?' The explanation must bridge the gap between the current state and the desired counterfactual state.

• Minimal Contrast: The closest possible counterfactual world where the decision is favorable defines the minimal change required.
• Contrastive Loss: Training objectives can be designed to minimize the distance between the factual instance and the generated counterfactual.
• User Comprehension: The delta between the current feature vector and the counterfactual must be translated into natural language, e.g., 'If your annual income were $5,000 higher, your application would be approved.'

Generating recourse should not require exposing sensitive training data or leaking information about other individuals in the dataset. The counterfactual explanation must be computed without violating data privacy.

• Differential Privacy: Noise can be added to the recourse generation process to prevent membership inference attacks.
• Model Inversion Defense: The counterfactual should not allow an adversary to reconstruct the decision boundary or training data.
• Local Explanations: Recourse should be derivable using only the model's query interface and the individual's own data, without requiring access to the full training set.

A causal definition of fairness where a decision is fair if it remains the same in a counterfactual world where the individual belonged to a different demographic group. This directly underpins recourse by modeling the minimal changes needed to flip a decision.

• Causal graphs distinguish discriminatory path-specific effects from legitimate influences
• Structural equation models generate the "what-if" scenarios necessary for actionable recourse
• Resolves the limitation of purely observational fairness metrics by modeling interventions

A model explainability technique that identifies the smallest set of input feature changes required to alter a model's prediction from an unfavorable outcome to a favorable one. This is the technical mechanism that makes algorithmic recourse operational.

• Generates actionable feature perturbations (e.g.,

A legal provision under Article 22 of the GDPR granting individuals the right to obtain meaningful information about the logic involved in automated decisions that produce legal or similarly significant effects. Algorithmic recourse operationalizes this right.

• Requires disclosure of decision-making logic and envisaged consequences
• The EU AI Act extends similar transparency mandates for high-risk systems
• Moves beyond passive transparency to actionable disclosure—explaining not just what happened, but how to remedy it

A design principle requiring that model explanations provide end-users with feasible, causal steps they can take to improve their outcome. This bridges the gap between passive interpretability and genuine algorithmic recourse.

• Feasibility constraints ensure recommendations respect immutable characteristics (age, ethnicity) and only suggest mutable features
• Causal reasoning prevents recommending changes that would not actually influence the decision
• Evaluated using metrics like recourse coverage—the percentage of negatively affected individuals who can achieve a favorable outcome through feasible actions

A systematic, independent evaluation of an algorithmic system to detect discriminatory outcomes using quantitative fairness metrics. Recourse analysis is increasingly integrated into bias audits to assess whether remediation pathways are equitably distributed.

• Audits measure disparities in recourse cost—whether some groups must make larger or more costly changes to reverse adverse decisions
• The Four-Fifths Rule and statistical parity tests are applied to recourse accessibility
• Mandated under New York City Local Law 144 for automated employment decision tools

A system design paradigm where a human operator provides active judgment and intervention for model outputs. In the context of recourse, HITL ensures that generated counterfactual recommendations are reviewed for real-world feasibility and ethical appropriateness before delivery to the affected individual.

• Prevents hallucinated recourse—suggestions that are mathematically valid but practically impossible
• Enables contextual judgment that purely algorithmic systems cannot replicate
• Required by the EU AI Act for high-risk automated decision systems under Article 14