Algorithmic Recourse

A recourse recommendation must involve features the individual can realistically change. It is useless to suggest an applicant be younger or change their historical credit defaults.

• Immutable features (age, birthplace) must be fixed or excluded from the counterfactual search.
• Actionable features (income, savings, debt ratio) define the feasible set of changes.
• Causal constraints ensure that changing one feature (e.g., education level) respects the real-world difficulty and downstream effects of that change.

The counterfactual explanation should change the minimum number of features necessary to flip the decision. A long list of required changes is cognitively overwhelming and practically infeasible.

• Objective: Minimize the L0 norm (count of changed features).
• Psychological load: Users are more likely to act on a concise plan (e.g., 'increase income by $5k') than a complex 10-step overhaul.
• Trade-off: Sparsity often competes with proximity; the closest counterfactual might require changing many features slightly, while a sparse one changes a few features significantly.

The counterfactual state should be as close as possible to the individual's current feature vector. Closer changes require less effort and are more credible.

• Distance metrics: Typically measured using L1 (Manhattan) or L2 (Euclidean) distance normalized by median absolute deviation.
• Plausibility: A counterfactual that is too far away may lie outside the true data manifold, representing an unrealistic or impossible combination of features.
• Contrast: Proximity ensures the explanation is local and faithful to the specific decision boundary affecting the user.

Effective recourse must respect the causal structure of the world. A naive counterfactual might suggest an intervention that breaks known causal laws or ignores downstream consequences.

• Structural Causal Models (SCMs): Used to model how changing one feature causally affects others.
• Downstream effects: Increasing 'credit card balance' to lower 'debt-to-income ratio' is a causal contradiction that a valid recourse engine must avoid.
• Average Causal Effect: The true cost of an action is measured by its total causal impact on all features, not just the direct change.

Providing a single counterfactual is insufficient. A diverse set of alternative explanations allows the user to choose the path that best fits their personal circumstances and preferences.

• Determinantal Point Processes (DPP): A mathematical mechanism used to select a subset of counterfactuals that are both valid and maximally different from each other.
• User agency: One user might prefer to increase income, while another might prefer to reduce debt; diverse options empower individual choice.
• Coverage: Diversity ensures the user sees fundamentally different ways to achieve the desired outcome, not just minor variations of the same plan.

The generated counterfactual must lie on the distribution of realistic data points. An explanation that suggests an impossible combination (e.g., a 16-year-old with a PhD) is not credible.

• Autoencoders & GANs: Generative models are used to constrain the search for counterfactuals to the learned data manifold.
• Negative feedback: Unrealistic explanations erode user trust in the entire system and may violate fairness principles by setting unattainable goals.
• In-distribution guarantee: Ensures the counterfactual is a plausible state that could actually be observed in the real world.

The primary technical mechanism for generating recourse. A counterfactual explanation identifies the minimal set of changes to an individual's input features that would have resulted in a favorable prediction.

• Answers the question: "What would need to be different for me to get the loan?"
• Relies on generative models or gradient-based optimization to find the nearest counterfactual point across the decision boundary.
• Key challenges include ensuring actionability (only mutable features are changed), sparsity (few features altered), and realism (the counterfactual lies on the data manifold).

A rigorous framework for distinguishing legitimate from discriminatory causal pathways. Recourse is meaningless if the features a person must change are actually proxies for protected attributes.

• Uses Structural Causal Models (SCMs) and directed acyclic graphs to map causal relationships.
• A fair recourse policy only suggests changes along non-discriminatory causal paths.
• Counterfactual fairness, a key causal definition, ensures a prediction would be identical in a world where only a protected attribute differed.

The set of rules that distinguish a useful recourse suggestion from an impossible one. An actionable explanation must respect real-world mutability.

• Immutable features: Race, birthplace, or age cannot be part of a recourse recommendation.
• Directional constraints: A bank cannot advise a customer to "decrease your age" or "reduce years of credit history."
• Causal constraints: Changing one feature (e.g., education level) may causally affect others (e.g., income), requiring causal reasoning to generate a coherent, multi-step recourse plan.

A broader explanatory paradigm that forms the psychological basis for recourse. Humans naturally ask "Why P, rather than Q?"

• A contrastive explanation highlights the difference between the actual outcome and a foil outcome.
• Algorithmic recourse is a specific, prescriptive form of contrastive explanation that not only identifies the difference but prescribes the actions to bridge the gap.
• Effective user interfaces present recourse as a clear contrast between the current state and the required target state.

The fairness principle that similar individuals should receive similar outcomes. This is the ethical foundation for why recourse must be consistent and achievable.

• Formalized by the Dwork et al. metric: the distance between two individuals' predictions should not exceed their distance in the input space.
• A system fails individual fairness if two nearly identical applicants receive opposite decisions with no clear, actionable difference.
• Recourse provides the path to similarity: it defines the precise distance an individual must travel in feature space to reach a favorable outcome.

A structured transparency document that should explicitly report a model's recourse capabilities and limitations. Standardized by Google, model cards are the vehicle for communicating recourse to auditors and users.

• A comprehensive model card details whether the system provides counterfactual explanations.
• It should report the average cost of recourse across different demographic groups to surface any disparities in the difficulty of achieving a positive outcome.
• Links to interactive recourse dashboards or APIs are often included.