Counterfactual Explanation

The non-negotiable requirement that the generated counterfactual instance must actually flip the model's prediction to the desired outcome. A counterfactual is only useful if it achieves its goal.

• Definition: The model's prediction for the counterfactual instance must be the target class (e.g., 'approved').
• Verification: This is a binary, objective check. If the counterfactual doesn't change the decision, it is invalid.
• Example: If a loan application was denied, a valid counterfactual must receive a 'granted' prediction from the same model.

The counterfactual should be as close as possible to the original instance in the feature space. This minimizes the cognitive load and the perceived effort required to change the outcome.

• Distance Metrics: Measured using L1 (Manhattan) or L2 (Euclidean) distance, weighted by feature importance or actionability.
• Sparsity: A subset of proximity, sparsity demands that the counterfactual changes the fewest number of features, not just the smallest total distance.
• Goal: To answer 'What is the easiest single thing I could have done differently?'

The counterfactual must only suggest changes to features that the end-user can realistically influence. Suggesting an immutable feature is useless and frustrating.

• Mutable Features: Variables a user can change (e.g., income, credit utilization, transaction amount).
• Immutable Features: Variables that are fixed (e.g., age, credit history length, past transaction timestamp).
• Constraint Enforcement: Algorithms must incorporate a mutable/immutable feature mask to ensure suggestions like 'be 5 years older' are never generated.

The counterfactual instance must lie within the distribution of realistic, feasible data points. It should not suggest a nonsensical combination of features that violates real-world logic.

• Data Manifold: The counterfactual must stay on the learned data manifold, avoiding outlier regions where no real data exists.
• Causal Constraints: It must respect known causal relationships (e.g., not suggesting 'increase savings balance' while simultaneously 'decreasing income').
• Density Checks: Often enforced by an autoencoder or a density-based penalty that rejects generated points in low-probability regions of the feature space.

Rather than a single counterfactual, a system should present multiple, meaningfully different paths to the desired outcome. This empowers the user with a choice of corrective strategies.

• Contrasting Explanations: Presenting one counterfactual that changes Feature A and another that changes Feature B.
• Determinantal Point Processes (DPP): A mathematical technique used to select a subset of counterfactuals that are both valid and maximally diverse from one another.
• User Benefit: A user might find one path easier or more palatable than another, increasing the utility of the explanation.

A robust counterfactual should reflect a causal relationship, not just a correlational one. It must answer what would have happened if the input had been different, based on a structural causal model.

• Structural Causal Models (SCM): The gold standard for generating causal counterfactuals, requiring a defined directed acyclic graph (DAG) of feature relationships.
• Interventional vs. Observational: A purely observational counterfactual might suggest 'increase credit score' by 'decreasing debt-to-income ratio,' but a causal model knows the latter is the true intervention point.
• Challenge: True causal counterfactuals are computationally harder to generate as they require a validated causal graph of the domain.

The ability to provide a human subject with a recommended set of actions to reverse an unfavorable algorithmic decision. While a counterfactual explanation identifies the minimal feature changes, algorithmic recourse ensures those changes are feasible and actionable in the real world.

• Feasibility constraints prevent suggesting impossible changes (e.g., 'be 10 years younger').
• Causal recourse goes beyond correlation to ensure the suggested action will actually cause the desired outcome flip.
• Directly addresses regulatory requirements for adverse action notices under the Fair Credit Reporting Act (FCRA).

A game-theoretic attribution method that decomposes a model's individual prediction into the additive contribution of each input feature. While SHAP explains why a decision was made, counterfactual explanations answer what would need to be different.

• Shapley values from cooperative game theory guarantee a fair distribution of credit across features.
• Local explainability pinpoints which features pushed a specific transaction into the fraud category.
• Often used as a preprocessing step to identify the most influential features before generating counterfactuals.

A legally mandated communication required under the Fair Credit Reporting Act (FCRA) and Equal Credit Opportunity Act (ECOA) when a consumer is denied credit or other benefits based on an automated decision. Counterfactual explanations provide the technical backbone for generating these notices.

• Must disclose the principal reasons for the denial.
• Counterfactuals translate opaque model logic into human-readable statements.
• Failure to provide accurate reasons exposes institutions to regulatory fines and litigation.

A specific algorithmic framework that generates explanations by identifying both pertinent positives (minimal features that must be present) and pertinent negatives (minimal features that must be absent) for a prediction. This directly operationalizes counterfactual reasoning.

• Pertinent positive: 'This transaction would be approved if the device fingerprint matched prior history.'
• Pertinent negative: 'This transaction would be approved if the amount were below $9,500.'
• Provides a more complete picture than a single counterfactual instance.

The institutional framework for identifying, assessing, and mitigating risks from model use. SR 11-7 guidance requires that model outputs be interpretable and that limitations are documented. Counterfactual explanations serve as a critical validation tool.

• Validators use counterfactuals to test if a model's decision boundary is conceptually sound.
• Helps identify proxy discrimination where protected class status is inferred through legitimate features.
• Essential for demonstrating model fairness during independent audits.

The statistical discipline of establishing cause-and-effect relationships from data, moving beyond mere correlation. Structural causal models (SCMs) enable the generation of counterfactuals that respect the causal mechanisms of the world.

• A causal counterfactual answers: 'What would the model predict if I intervened to change X?'
• Avoids suggesting spurious changes that correlate with approval but don't cause it.
• Do-calculus provides the mathematical framework for reasoning about interventions.