Actionable Recourse

The action set formally specifies the permissible modifications a user can make to each feature. It defines the boundary between actionable and non-actionable changes.

• Mutable features: Income, savings rate, loan term
• Immutable features: Age, place of birth, historical defaults
• Directional constraints: Income can only increase, age cannot decrease

Without a defined action set, a model might recommend changing your birth year to qualify for a loan—mathematically valid, practically useless.

Immutable features are protected attributes that must be held constant during counterfactual generation. These include:

• Demographic constants: Age, birthplace, ethnicity
• Historical facts: Prior convictions, past loan defaults
• Genetic markers: Inherited health conditions

Algorithms enforce immutability through hard constraints that zero out gradients for these features, ensuring the optimization process never suggests changing them.

Feasibility constraints encode real-world limitations beyond simple mutability. They ensure recommendations respect causal relationships and practical boundaries.

• Monotonicity: Education level can only increase, never decrease
• Causal consistency: Changing 'number of dependents' should not precede 'marital status' changes
• Range bounds: Credit score recommendations must stay within valid ranges (300-850)

These constraints prevent the generation of implausible counterfactuals that, while technically actionable, violate real-world logic.

Recourse robustness measures whether a recommended action remains valid after the underlying model is retrained or slightly updated. A fragile recourse recommendation becomes invalid with minor model shifts.

• Robust recourse survives model retraining on new data
• Fragile recourse exploits transient decision boundary quirks
• Techniques include adversarial training and uncertainty-aware optimization

Without robustness, a user might follow a recommendation only to find the model has changed and the action no longer works.

Causal recourse uses a Structural Causal Model (SCM) to ensure recommended changes respect cause-and-effect relationships. Unlike purely statistical approaches, causal recourse accounts for downstream effects.

• Changing 'education level' may causally affect 'income'
• Interventions are modeled using do-calculus operators
• Prevents recommending changes that would have contradictory ripple effects

This approach ensures that acting on one recommendation doesn't inadvertently invalidate another feature's value.

Providing diverse counterfactuals gives users multiple distinct paths to achieve a desired outcome. A single recommendation may be infeasible for a specific individual.

• Financial example: Increase income by $10k OR reduce debt by $15k OR extend loan term by 5 years
• Diversity metrics measure how different the proposed paths are
• Avoids recourse dead ends where the only option is practically impossible

Diversity transforms recourse from a single prescription into a menu of actionable options.