Counterfactual Explanations

The primary goal of a counterfactual explanation is to provide a feasible path for a user to achieve a desired outcome. An actionable counterfactual suggests changes that are within the user's control and are realistic to implement.

• Example: For a loan denial, an actionable counterfactual might be: "Increase your annual income by $5,000." A non-actionable one would be: "Be 10 years older."
• This characteristic is central to explainability score validation, as an explanation's usefulness is tied to its practical guidance.

A high-quality counterfactual should be as close as possible to the original input instance. This is typically measured using a distance metric (e.g., L1 or L2 norm) in the feature space. The explanation answers: "What is the smallest change needed?"

• Sparse changes are preferred, altering the fewest number of features.
• Proximity ensures the suggested alternative is relevant and comparable to the original case, not an entirely different data point.
• This property is quantitatively assessed in faithfulness score validation to ensure the explanation reflects the model's local decision boundary.

The generated counterfactual must be valid—meaning it leads the model to output the desired prediction (e.g., changes a 'deny' to an 'approve'). It must also be plausible, representing a realistic data point that could exist in the real world.

• Implausible Example: A counterfactual suggesting a 2-meter-tall person weighs 20kg violates physical laws.
• Plausibility is enforced through constraints on feature relationships and data manifold proximity. This is a key focus of synthetic data fidelity assessment when generating counterfactuals.
• A valid, plausible counterfactual passes a basic simulatability test: a human can understand the change and believe it would alter the outcome.

Counterfactuals must respect causal relationships and immutable features. You cannot suggest changing a person's birthplace or age. A robust method incorporates domain knowledge to avoid nonsensical suggestions.

• Immutable Features: Race, gender, past events.
• Causal Dependencies: Increasing 'years of education' may causally influence 'income'; they cannot be changed independently without violating realism.
• Ignoring causality can lead to unfaithful explanations that the model would not actually follow. This connects to perturbation analysis for validating that suggested changes align with the model's learned patterns.

For a given instance, there are often multiple valid counterfactuals. A good explanation system should be able to generate a diverse set of alternative paths to the desired outcome, providing users with choice.

• Example for a loan denial: Option A: "Increase income by $5k." Option B: "Reduce debt by $2k."
• Diversity prevents over-reliance on a single, potentially suboptimal path and helps users find the most actionable route for their circumstances.
• Evaluating diversity is part of explanation robustness assessment, ensuring the method doesn't collapse to a single, brittle solution.

Counterfactual explanations are inherently contrastive. They do not explain why the current outcome was reached in absolute terms, but why it was reached instead of a specific alternative outcome. They answer: "Why was I denied a loan, rather than approved?"

• This aligns with human reasoning, which often seeks contrasting cases to understand causality.
• The contrastive explanation is defined by the desired class (the 'counterfactual' class).
• This characteristic makes them particularly useful for recourse and debugging, as they focus on the delta between outcomes.

Proximity measures the distance between the original input and the generated counterfactual. A valid counterfactual should be minimally distant, representing the smallest realistic change to alter the prediction. Common distance metrics include:

• L1 (Manhattan) or L2 (Euclidean) distance for continuous features.
• Hamming distance or custom categorical distance for discrete features.
• Weighted distances that account for feature-specific plausibility or cost.

Low proximity indicates the explanation suggests unrealistic or drastic changes, reducing its practical utility.

Sparsity quantifies how many input features were changed to generate the counterfactual. A sparse explanation, where only 1-2 key features are altered, is more interpretable and actionable than one requiring changes across many dimensions. Evaluation involves:

• Counting the number of features with non-zero change magnitude.
• Assessing if changed features are actionable (e.g., income) versus immutable (e.g., age).
• Optimizing for feature-change sparsity as a primary objective during counterfactual generation.

High sparsity aligns with the principle of parsimony, aiding in root-cause analysis.

Validity is a binary metric confirming the counterfactual input actually produces the desired target prediction from the model. It is the most fundamental requirement. Evaluation is straightforward:

• Pass the generated counterfactual through the original model.
• Check if the model's output matches the specified contrastive class (e.g., 'loan approved' instead of 'denied').

A failure of validity indicates the explanation method is not faithful to the model's decision boundary.

Plausibility assesses whether the counterfactual example is realistic and could exist in the real world. An implausible counterfactual (e.g., 'change age from 25 to -5') is not actionable. Evaluation methods include:

• Measuring distance to the training data manifold using k-NN or density estimators.
• Using autoencoder reconstruction error; low error indicates the point lies on the learned data distribution.
• Applying domain constraints (e.g., age > 0, systolic BP > diastolic BP) as hard feasibility checks.

This metric guards against adversarial examples that flip the prediction but are nonsensical.

For a given instance, there are often multiple valid counterfactual paths. Diversity evaluates a set of counterfactuals to ensure they propose meaningfully different alternative scenarios. This is crucial for providing users with options. It is measured by:

• Calculating pairwise distance (e.g., L2) between counterfactuals in the set.
• Ensuring features changed vary across the set (e.g., one suggests increasing income, another suggests reducing debt).
• Avoiding mode collapse where all generated counterfactuals are nearly identical.

High diversity supports exploratory analysis and robust decision-making.

The most advanced evaluation considers known causal relationships between features. A counterfactual suggesting 'increase education level' to get a loan may be invalid if education level causally influences income—changing one without the other may be unrealistic. Evaluation involves:

• Integrating a causal graph (DAG) to check if proposed changes respect causal dependencies.
• Distinguishing actionable features (e.g., savings) from non-actionable (e.g., past diagnosis) or immutable ones (e.g., race).
• Assessing if the explanation suggests realistic interventions within a feasible timeframe.

This moves evaluation from statistical proximity to real-world feasibility.

A unified framework based on cooperative game theory that assigns each feature an importance value for a specific prediction. Unlike counterfactuals, SHAP provides a global and local attribution score, explaining the contribution of each feature to the deviation from a baseline expected output. It is computationally intensive but provides a theoretically grounded distribution of 'credit' among features.

A model-agnostic technique that approximates a complex model locally around a single prediction with a simple, interpretable surrogate model (like linear regression). It answers 'what features were important for this prediction?' by perturbing the input and observing output changes. While LIME highlights influential features, a counterfactual explicitly defines the minimal changes needed to flip the prediction.

A core validation metric that quantifies how accurately an explanation reflects the true reasoning process of the underlying model. For counterfactuals, faithfulness can be measured by:

• Implementing the suggested changes and verifying the prediction flips as expected.
• Using perturbation analysis to see if other, similar changes produce the same outcome. A high faithfulness score indicates the counterfactual is a credible 'what-if' scenario for the model.

Explanations that answer the question 'Why P rather than Q?' They explicitly contrast the actual prediction (P) with a plausible alternative (Q). A counterfactual explanation is a specific type of contrastive explanation where Q is the desired outcome. The explanation highlights the minimal features that, if changed, would lead to Q instead of P, making the contrast inherent and explicit.

A family of techniques for generating or validating explanations by systematically modifying input features and observing the impact on the model's output. It is the foundational mechanism behind many explanation methods:

• Generating Counterfactuals: Searching the feature space via perturbations to find the minimal change that alters the prediction.
• Validating Explanations: Testing if perturbing features deemed important by an explanation (e.g., SHAP) causes significant output change, assessing its infidelity.

A model-agnostic explanation method that provides a high-precision rule (the 'anchor') sufficient to guarantee a prediction locally. An anchor is a set of if-then conditions on features (e.g., 'IF Age > 50 AND Blood_Pressure = High'). While an anchor explains what secures the current prediction, a counterfactual explains what would change it. They are complementary: anchors describe a robust region for the prediction; counterfactuals describe the exit from that region.