Mahalanobis Distance

Unlike Euclidean distance, which treats all features as independent and equally scaled, the Mahalanobis distance uses the inverse of the covariance matrix to rescale the feature space.

• Unitless: Converts all features to a common, dimensionless scale.
• Correlation Correction: If two features are highly correlated, the metric stretches the space along the axis of correlation, making distances along that axis smaller.
• Formula: $D_M(x, \mu) = \sqrt{(x - \mu)^T \Sigma^{-1} (x - \mu)}$, where $\Sigma$ is the covariance matrix.

Using Mahalanobis distance as a loss term during counterfactual generation ensures the resulting instance lies within the high-density region of the training data.

• Avoids Adversarial Artifacts: Euclidean distance can produce counterfactuals that are close in L2 norm but violate feature correlations (e.g., suggesting a 50-year-old with 2 years of credit history).
• Distributional Proximity: Minimizing the Mahalanobis distance to the target class centroid or the full distribution pushes the counterfactual toward realistic, plausible feature combinations.

Computing the Mahalanobis distance is mathematically equivalent to applying a whitening transformation to the data and then measuring standard Euclidean distance.

• Whitening Process: The data is transformed by $\Sigma^{-1/2}$, which decorrelates the features and scales each to unit variance.
• Computational Insight: This equivalence allows engineers to use standard Euclidean nearest-neighbor algorithms on the transformed space to find Mahalanobis-nearest counterfactuals.

The standard sample covariance matrix becomes ill-conditioned or singular when the number of features $p$ approaches the number of samples $n$.

• Minimum Covariance Determinant (MCD): A robust estimator that finds the subset of $h$ observations with the smallest covariance determinant, ignoring outliers.
• Shrinkage Estimators: Techniques like Ledoit-Wolf shrinkage combine the sample covariance with a structured target (e.g., identity matrix) to produce a well-conditioned estimate: $\Sigma_{shrunk} = (1-\alpha)\Sigma_{sample} + \alpha T$.

Squared Mahalanobis distances of points drawn from a multivariate Gaussian distribution follow a chi-squared distribution with $p$ degrees of freedom.

• Statistical Thresholding: A point can be flagged as an outlier if $D_M^2(x) > \chi^2_{p}(0.975)$, providing a principled, distribution-aware cutoff.
• Counterfactual Validation: This property is used to verify that a generated counterfactual is not an outlier in the target class distribution, ensuring plausibility.

The Mahalanobis distance is a specific case of a generalized quadratic distance metric, where the matrix $M$ in $d_M(x, y) = (x-y)^T M (x-y)$ is the inverse covariance matrix $\Sigma^{-1}$.

• Large Margin Nearest Neighbor (LMNN): A metric learning algorithm that learns an optimal $M$ to minimize the distance between points of the same class while enforcing a margin for different classes.
• Application: A learned Mahalanobis metric can be tailored specifically for counterfactual search, prioritizing dimensions that are most relevant to the decision boundary.

A counterfactual instance that lies within the high-density region of the training data distribution. While Euclidean distance might generate a close but unrealistic point, Mahalanobis Distance directly optimizes for plausibility by penalizing deviations that ignore feature correlations.

• Ensures the explanation is realistic, not an adversarial artifact
• Respects the covariance structure of the data
• Prevents suggesting impossible feature combinations

A metric quantifying the distance between the original input instance and the generated counterfactual. While L1 and L2 norms are common, they treat features as independent. Mahalanobis Distance provides a more sophisticated proximity measure by scaling axes according to variance and rotating them to decorrelate features.

• L1: Sum of absolute differences
• L2: Straight-line Euclidean distance
• Mahalanobis: Correlation-adjusted statistical distance

A hard rule encoded into a counterfactual generation algorithm that prevents the modification of immutable features or enforces causal monotonicity. When using Mahalanobis Distance, feasibility constraints are layered on top of the distance minimization to ensure the path to the counterfactual respects real-world logic.

• Prevents changes to immutable features like age
• Enforces causal directionality
• Combined with Mahalanobis for actionable, plausible recourse

A subset of algorithmic recourse that constrains recommended changes to only those features an individual can realistically control. Mahalanobis Distance is critical here because it prevents the algorithm from exploiting correlated but non-actionable proxies to achieve a minimal distance flip.

• Distinguishes between statistical and actionable proximity
• Prevents 'cheating' via immutable feature correlations
• Ensures the recommended path is under the user's control

A set of multiple, distinct counterfactual instances providing alternative paths to a desired outcome. Algorithms using Mahalanobis Distance can generate diverse explanations by exploring different directions within the normalized, decorrelated feature space, ensuring variety without sacrificing plausibility.

• Avoids a single, potentially infeasible recommendation
• Explores the Mahalanobis ellipsoid for distinct valid points
• Provides users with meaningful choice in recourse paths

A formal framework representing variables and their causal dependencies through structural equations. While Mahalanobis Distance captures observational correlations, an SCM captures causal mechanisms. Advanced counterfactual generators combine both: using the SCM to constrain changes and Mahalanobis distance to measure proximity in the causally-aware feature space.

• Encodes cause-effect relationships
• Enables interventional and counterfactual queries
• Complements Mahalanobis for causally valid explanations