Counterfactual Explanation

The most fundamental requirement: the generated counterfactual instance must successfully flip the model's prediction to the predefined target outcome. This is a binary pass/fail metric.

• A counterfactual that does not change the prediction is not an explanation—it is a failed search.
• Validity is verified by passing the counterfactual through the model and checking if f(x_cf) == y_target.
• In multi-class settings, this means the target class probability must exceed all others.
• Algorithms like Growing Spheres guarantee validity by design, stopping only when the decision boundary is crossed.

A valid counterfactual should be as close as possible to the original input instance. Proximity operationalizes the principle of minimal change.

• Measured using distance functions such as L1 (Manhattan) or L2 (Euclidean) norms.
• A lower distance means fewer and smaller feature perturbations, making the explanation easier to interpret.
• The Mahalanobis Distance is often preferred because it accounts for feature correlations and variances in the data distribution.
• Trade-off: overly strict proximity constraints can make it impossible to find a valid counterfactual.

A sparse counterfactual changes only a small number of features, even if the total distance is not minimal. Humans struggle to process high-dimensional changes.

• Sparse explanations are cognitively tractable: 'Change your income and loan amount' is clearer than 'Adjust 47 variables.'
• Enforced via L0 regularization or by explicitly limiting the number of altered features during generation.
• Sparse counterfactuals directly support algorithmic recourse by highlighting the few levers a user can pull.
• Example: A loan rejection counterfactual that only changes 'credit score' and 'debt-to-income ratio' is more actionable than one altering a dozen fields.

A plausible counterfactual lies within the high-density region of the training data distribution. It must represent a realistic, achievable instance.

• Counterfactuals in low-density or empty regions of feature space are adversarial artifacts, not genuine explanations.
• An implausible example: suggesting a person be 20 years old with 30 years of work experience.
• Plausibility is enforced by adding a density constraint or using autoencoders to ensure the counterfactual decodes to a realistic sample.
• This property is critical for actionable recourse—a recommendation must be physically possible in the real world.

Rather than a single counterfactual, a system should generate multiple, distinct alternatives that all achieve the desired outcome.

• Different users have different constraints; one path to recourse may be infeasible while another is trivial.
• Diversity is enforced by adding a determinantal point process (DPP) loss or simply penalizing similarity between generated instances.
• Example: For a denied credit application, one counterfactual might suggest increasing income, while another suggests reducing existing debt.
• Diverse counterfactuals empower user choice and reveal the model's decision boundary from multiple angles.

A counterfactual must respect the causal structure of the world. Changing a feature should not violate known cause-effect relationships.

• If 'education level' is changed, 'age' must also increase to maintain consistency with causal dependencies.
• Immutable features like birthplace or race must be held constant; changing them produces a nonsensical explanation.
• Enforced by integrating a Structural Causal Model (SCM) into the generation process.
• Without causal constraints, counterfactuals can recommend impossible states—undermining trust and violating counterfactual fairness principles.

The actionable bridge between an explanation and a real-world outcome. While a counterfactual explanation identifies the minimal change to flip a prediction, algorithmic recourse translates that change into a set of steps an end-user can actually take. It shifts the focus from 'why was I denied?' to 'what can I do to be approved?'

The mathematical engine for true counterfactual reasoning. An SCM defines a system of structural equations (X_i = f_i(PA_i, U_i)) that represent causal mechanisms, not just correlations. This allows you to compute answers to interventional and counterfactual queries using the three-step process of abduction, action, and prediction. Without an SCM, a 'counterfactual' is merely a perturbed data point.

A rigorous, individual-level fairness criterion. A predictor (\hat{Y}) is counterfactually fair if its output for an individual is identical in the actual world and a counterfactual world where a sensitive attribute (A) (e.g., race, gender) had been different. This leverages SCMs to define fairness by comparing an individual to themselves under a hypothetical change, not to a group average.

A counterfactual must not only be close but also realistic. A plausible counterfactual lies within the high-density region of the training data manifold. This prevents generating adversarial or nonsensical instances—like a person with an age of -5 or a credit history of 200 years—that cross the decision boundary but violate real-world logic. Measured using Mahalanobis distance or density estimation.

A multi-objective framework for benchmarking generation algorithms. Key metrics include:

• Validity: Did the prediction flip?
• Proximity: How far is the counterfactual from the original?
• Sparsity: How few features were changed?
• Diversity: Are multiple distinct paths offered?
• Plausibility: Is the instance realistic? No single metric is sufficient; a good explanation balances all five.

The cognitive science foundation of human explanation. A contrastive explanation answers the question 'Why P and not Q?' by highlighting the minimal set of features that differentiate the factual outcome from a presented foil. In machine learning, this is operationalized by generating a counterfactual instance Q and explaining the difference between the original input and Q. It mirrors how humans naturally seek explanations.