Algorithmic Recourse

The recourse provided must only suggest changes to features an individual can realistically control. Recommending an applicant decrease their age or change their marital status is not actionable.

• Immutable Features: Features like age, ethnicity, or place of birth must be fixed or excluded from the recourse generation.
• Actionable Features: Suggestions should focus on mutable inputs like income, savings rate, credit utilization, or specific skill certifications.
• Constraint Handling: Effective systems incorporate hard constraints, ensuring the path to a positive decision does not violate real-world logic or legal boundaries.

Recourse must be based on causal relationships, not mere correlations. A correlational model might suggest opening a new credit card to improve a credit score, but the causal effect of a hard inquiry could temporarily lower it.

• Structural Causal Models (SCMs): These models map the cause-and-effect relationships between variables, allowing the system to predict the true downstream impact of a change.
• Intervention vs. Observation: The system must answer 'What will happen if I do X?' rather than 'What do people who look like X typically have?'
• Avoiding Spurious Correlations: Causal reasoning prevents the system from suggesting nonsensical actions based on coincidental patterns in historical data.

A recourse path should remain valid even if the underlying model is slightly updated or retrained. A fragile explanation that becomes invalid after a minor model refresh erodes user trust.

• Model Shift Resilience: The suggested action should flip the decision for a set of plausible nearby models, not just the exact current one.
• Adversarial Robustness: The path should not be easily invalidated by small, random perturbations to the individual's other features.
• Temporal Stability: The explanation should hold true long enough for the individual to realistically implement the recommended changes.

The path to a favorable outcome must account for the real-world difficulty and cost of making a change. Increasing annual income by $50,000 is technically actionable but may be infeasible for most individuals.

• Cost Functions: A mathematical function quantifies the effort, time, or monetary cost required to change a specific feature.
• Personalized Difficulty: The system should ideally learn individual-specific cost functions, recognizing that 'improve education level' has a vastly different cost for different people.
• Diverse Paths: Providing multiple recourse options with varying cost profiles allows the individual to choose the path that best fits their circumstances.

An effective explanation should ask the individual to change as few features as possible. A long list of required changes is overwhelming and practically impossible to implement.

• Minimal Intervention Sets: The algorithm should search for the smallest set of feature changes that flips the decision.
• Cognitive Load: A sparse explanation is easier for a human to understand, remember, and act upon.
• Contrastive Nature: Sparse recourse naturally forms a contrastive explanation: 'You were denied. If your feature X and Y were at these values, you would have been approved.'

The process of generating recourse should not leak sensitive information about other individuals in the training data. A counterfactual explanation like 'If you lived in this specific affluent neighborhood, you'd be approved' can implicitly reveal private demographic patterns.

• Differential Privacy for Recourse: Techniques exist to generate counterfactuals that are statistically indistinguishable from those generated without any single individual's data.
• Avoiding Attribute Inference: The explanation itself must not allow the user to infer the protected attributes of others who received a favorable outcome.
• Model-Agnostic Privacy: The recourse method should protect the privacy of the training data regardless of the underlying model type.

Robust algorithmic recourse requires causal reasoning, not just correlation. Without understanding the causal structure of the data, counterfactual explanations can suggest changes that are impossible or produce unintended downstream effects. For instance, suggesting a user 'increase their education level' to qualify for a loan ignores that education causally affects income, which also affects the decision.

• Structural Causal Models (SCMs): Formal frameworks for encoding causal relationships between variables
• Intervention vs. observation: Recourse requires estimating the effect of doing something, not just observing correlations
• Causal fairness: Definitions like counterfactual fairness explicitly use causal models to define non-discrimination

Not all model inputs are mutable. A critical step in building recourse systems is classifying features by their actionability and mutability:

• Mutable and actionable: Credit card balance, savings rate, application timing—features an individual can directly change
• Mutable but non-actionable: Age, zip code—features that change over time but cannot be deliberately altered
• Immutable: Birth country, genetic markers—features fixed at the time of decision

Recourse algorithms must constrain counterfactual generation to only suggest changes in the first category, respecting both real-world feasibility and ethical boundaries around asking individuals to change protected characteristics.

Even well-intentioned recourse systems can perpetuate inequity. Research shows that the cost of recourse—the difficulty of achieving a favorable outcome—can vary systematically across demographic groups due to historical disparities in feature distributions.

• Equal recourse effort: A fairness criterion requiring that the average cost to achieve recourse is similar across protected groups
• Disparate recourse availability: Situations where some groups have no feasible path to a positive decision within realistic constraints
• Robust recourse: Ensuring counterfactual explanations remain valid after model updates, preventing 'recourse that expires'

The broader family of techniques that produce explanations by contrasting the actual outcome with a hypothetical alternative. Beyond counterfactuals for recourse, this includes:

• Contrastive explanations (CE): Identifying a minimal set of features that, if absent, would change the classification—focuses on the present state rather than future changes
• Pertinent positives and negatives: Explaining a decision by showing both what must be present (positives) and what must be absent (negatives) for the decision to hold
• Anchor explanations: High-precision if-then rules that 'anchor' a prediction locally, providing sufficient conditions for the decision

Documentation frameworks that communicate a model's intended use, limitations, and fairness properties—including whether recourse mechanisms are available. Model Cards serve as the governance layer that connects technical recourse implementations to organizational accountability.

• Intended use: Specifies the populations and contexts for which recourse paths have been validated
• Ethical considerations: Discloses known disparities in recourse cost or availability across groups
• Evaluation results: Reports quantitative metrics on counterfactual validity, proximity, and diversity

Standardized transparency artifacts ensure that recourse is not just technically possible but organizationally visible and auditable.