Counterfactual Explanations

The core objective is to find the closest possible world to the original input that results in a different outcome. This is often framed as an optimization problem minimizing the L1 or L2 distance between the original feature vector and the counterfactual state. The goal is to identify the smallest, most actionable adjustment—such as 'increase salary by $5,000' rather than 'change job sector and relocate'—to provide realistic recourse.

Unlike standard feature attribution, true counterfactuals require a Structural Causal Model (SCM) to ensure changes respect real-world dependencies. A purely associational counterfactual might suggest 'reduce age by 10 years' to flip a loan decision, which is impossible. A causal approach enforces intervention consistency, ensuring that changing a parent variable correctly propagates to its children, preventing unrealistic or unactionable explanations.

Generating a single counterfactual is often insufficient. Frameworks like DiCE (Diverse Counterfactual Explanations) optimize for a set of explanations that are both diverse and feasible. Diversity ensures the user sees multiple distinct paths to recourse (e.g., 'reduce debt-to-income ratio' or 'increase down payment'). Feasibility constraints, often learned from data density, prevent suggestions that fall outside the plausible data manifold.

A valid counterfactual must respect actionable features—variables the end-user can actually influence. It is critical to hard-code immutable attributes like race, birthplace, or genetic markers as non-modifiable during the search process. Generating a counterfactual that suggests 'change gender' to receive a loan approval is not only useless but constitutes evidence of illegal discriminatory logic within the black-box model.

Counterfactual explanations directly fulfill the 'right to explanation' mandated by regulations like the GDPR and the EU AI Act. They provide a contrastive, human-friendly format that is psychologically intuitive. Auditors use them to test for algorithmic fairness by checking if similarly situated individuals from different protected groups require vastly different degrees of change to achieve the same positive outcome, revealing potential proxy discrimination.

A direct extension of standard counterfactuals that generates multiple distinct scenarios rather than a single minimal change. DiCE optimizes for both proximity to the original input and diversity among the generated counterfactuals, ensuring users see a range of actionable options—such as 'increase income by $5K' or 'reduce credit utilization by 15%'—rather than a single, potentially impractical path to a desired outcome.

Standard Shapley values measure statistical association, which can produce misleading importance scores when features are correlated. Causal Shapley values incorporate a Structural Causal Model (SCM) of the data-generating process to assign importance based on causal effects rather than mere correlations. This aligns closely with counterfactual reasoning by answering: 'How much did this feature cause the prediction to change?'

While counterfactuals identify the minimal change to flip a prediction, anchors identify the minimal set of conditions that guarantee a prediction stays the same. An anchor is a high-precision rule—e.g., 'education > Bachelor's AND hours-per-week > 30'—such that no matter how other features vary, the model's prediction remains unchanged. Anchors provide sufficiency conditions, complementing counterfactuals' necessity logic.

Counterfactuals ask 'what if my features were different?' Influence functions ask 'what if a specific training example were removed?' Adapted from robust statistics, this technique traces a model's prediction back to its training data by estimating how upweighting or removing a particular data point would change the loss. This enables data-level counterfactual reasoning for debugging model behavior and identifying mislabeled examples.

The formal mathematical framework underlying all counterfactual reasoning. An SCM represents variables and their causal relationships using structural equations and a directed acyclic graph. Unlike purely associational models, SCMs support three layers of causal reasoning:

• Observational: What is?
• Interventional: What if I do?
• Counterfactual: What if I had done otherwise? Counterfactual explanations require at minimum a partial causal model to ensure the proposed changes are actionable and realistic.

A distribution-free framework that produces prediction sets with rigorous coverage guarantees. While counterfactuals explain why a decision was made, conformal prediction quantifies how certain that decision is. For any given confidence level (e.g., 90%), conformal prediction generates a set of possible labels guaranteed to contain the true label with that probability. This provides the uncertainty quantification layer that counterfactual explanations alone lack.