Plausible Counterfactual

The generated instance must reside in a region of the feature space where the training data probability density is high. This prevents the algorithm from suggesting changes that, while mathematically minimal, produce a nonsensical or impossible data point. Techniques often enforce this by using autoencoders to project perturbations back onto the learned data manifold or by minimizing the Mahalanobis distance to the data centroid rather than simple Euclidean distance. A classic failure mode is a counterfactual for loan approval that suggests a negative age to flip the decision.

Plausible counterfactuals respect the causal structure of the world. Changing a feature must not violate known causal dependencies. For instance, increasing 'education level' should not decrease 'age'. This is enforced by integrating a Structural Causal Model (SCM) or a causal graph into the generation process. The algorithm must compute the downstream effects of an intervention rather than treating all features as independently mutable. Without causal constraints, a counterfactual might suggest a user change their 'years of credit history' without acknowledging this is causally linked to their age.

A counterfactual is only plausible if the recommended changes are within the user's action set. An explanation that tells a loan applicant to reduce their age by 10 years is not actionable. Plausibility requires hard constraints that freeze immutable features (age, birthplace) and respect monotonic relationships. The action set formally defines the permissible modifications for each feature, distinguishing between mutable and immutable attributes. This directly links plausibility to the concept of algorithmic recourse, ensuring the explanation empowers the user rather than frustrating them.

The counterfactual must preserve the semantic meaning of categorical and ordinal features. A plausible counterfactual for a mortgage application would not suggest changing the 'property type' from 'single-family home' to 'commercial warehouse' if the loan product is strictly for residential properties. This requires the generation algorithm to operate within valid categorical ranges and respect ordinal relationships. For example, 'education level' can only increase or decrease along its defined ordinal scale, not jump to an undefined category. This is often enforced through one-hot encoding constraints and ordinal regularization.

Standard L1 or L2 distance metrics are insufficient for measuring plausibility because they treat all feature space regions equally. Plausible counterfactuals require distance metrics that account for the underlying data distribution. The Mahalanobis distance weights perturbations by the inverse covariance matrix of the training data, penalizing changes that move along low-variance, correlated feature directions. Alternatively, kernel density estimation can be used to directly score the likelihood of a generated point under the training distribution, rejecting candidates that fall below a density threshold.

A plausible counterfactual should remain valid even after the underlying model is updated with new data. If a generated explanation flips the prediction only by exploiting a brittle, transient artifact of the current decision boundary, it lacks recourse robustness. Plausibility is enhanced by generating counterfactuals that cross the decision boundary with a margin, landing deep inside the target class's high-density region rather than on the boundary itself. This ensures the explanation remains actionable and valid over time, which is critical for regulatory compliance in financial services.

A distance metric that accounts for feature correlations and variances in the training data distribution. Unlike Euclidean distance, it measures distance in units of standard deviation along the principal components of the data.

• Captures the shape of the data distribution
• A small Mahalanobis distance indicates the counterfactual lies in a high-density region
• Essential for generating plausible rather than merely proximal counterfactuals

A formal framework representing variables and their causal dependencies through structural equations. SCMs enable the computation of interventional and counterfactual queries.

• Encodes domain knowledge about cause-effect relationships
• Ensures counterfactuals respect causal constraints (e.g., changing 'education' may affect 'income')
• Distinguishes between correlation and causation to prevent unrealistic explanations

A hard rule encoded into a counterfactual generation algorithm that prevents the modification of immutable features or enforces causal monotonicity.

• Example: Age cannot decrease; it is an immutable feature
• Example: Increasing education level should not decrease years of work experience
• These constraints define the boundary between a plausible explanation and an adversarial artifact

A metric quantifying the distance between the original input instance and the generated counterfactual. Often measured using L1 or L2 norms.

• Enforces the principle of minimal change
• A counterfactual that is too distant may be implausible or costly to achieve
• Must be balanced with plausibility—the closest counterfactual may lie outside the data manifold

A subset of algorithmic recourse that constrains recommended changes to only those features an individual can realistically control or modify in the real world.

• Distinguishes between 'change your age' (impossible) and 'increase your savings' (actionable)
• Requires an action set specifying permissible modifications
• Directly addresses the practical utility of counterfactual explanations for end-users

A set of multiple, distinct counterfactual instances that provide a user with alternative paths to achieve a desired outcome.

• Avoids presenting a single, potentially infeasible recommendation
• Example: A loan applicant could be shown 'increase income by $5k' OR 'reduce debt by $10k'
• Improves recourse robustness by offering fallback options