Counterfactual Explanations

An actionable counterfactual suggests changes to features that the individual or system can realistically influence. For example, 'Increase your credit score by 30 points' is actionable, while 'Be 10 years younger' is not. In a knowledge graph, this property is enforced by ontological constraints that define mutable attributes (e.g., hasSkill) versus immutable ones (e.g., dateOfBirth).

This is the core defining property. A valid counterfactual must describe the smallest possible change to the input that results in a different (usually desired) prediction. It answers: "What is the least I need to change?"

• Measured by distance metrics like L1 (Manhattan) or L2 (Euclidean) in feature space.
• In graph-structured data, proximity is measured in graph edits: the minimal number of added/removed edges or changed node attributes required to flip the prediction.

A plausible counterfactual must describe a realistic, data-supported alternative. It should lie within the observed data distribution, not in an impossible region of feature space. For instance, 'Earn a PhD in 1 month' is implausible.

• Knowledge graphs enforce plausibility through ontological rules and relational constraints (e.g., a Person node cannot have a manufacturedIn property).
• Methods often use graph autoencoders or check against known factual triples to ensure generated counterfactuals are semantically valid.

The suggested changes should respect known or inferred cause-and-effect relationships. Changing one feature (e.g., annualIncome) may necessitate changes in others (e.g., taxBracket). Ignoring causality can lead to nonsensical or ineffective recommendations.

• Knowledge graphs with causal edges (e.g., causes or influences) provide a structured model to generate causally coherent counterfactuals.
• This property bridges counterfactual explanations with Structural Causal Models (SCMs).

For a single instance, there may be multiple, equally valid paths to a desired outcome. Providing a diverse set of counterfactuals offers users meaningful choice. For a loan denial, diverse explanations could be:

• Increase income by $5,000.
• Reduce debt-to-income ratio by 10%.
• Add 2 years of employment history.

Algorithms like Diverse Counterfactual Explanations (DiCE) are explicitly designed to optimize for this property.

Closely related to proximity, sparsity specifically demands that the explanation alters as few features as possible. Humans find sparse explanations (changing 1-2 things) more interpretable than those requiring changes to many features. In a graph neural network (GNN) context, a sparse explanation might identify a single decisive subgraph or a minimal set of node feature perturbations, rather than diffusing importance across the entire input graph.

Engineers use counterfactuals to stress-test and debug machine learning models by identifying decision boundaries and sensitivity points. For example, generating counterfactuals can reveal if a loan approval model is overly sensitive to a single feature like postal code, indicating potential bias or a data leakage issue. This process validates that the model's logic aligns with domain expertise encoded in a knowledge graph.

• Identify edge cases and model fragility.
• Verify feature robustness against small, realistic perturbations.
• Contrast model logic with business rules or causal graphs.

This use case provides individuals with actionable steps to achieve a desired outcome from an automated system. Unlike a simple denial, algorithmic recourse answers "what can I do?" For instance, a system might tell a loan applicant: "Your application would be approved if your credit score increased by 20 points."

• Promotes fairness by enabling self-improvement.
• Shifts from static decisions to dynamic guidance.
• Requires feasible, actionable changes grounded in reality (e.g., via knowledge graph constraints).

Counterfactuals help organizations comply with regulations like the EU AI Act and GDPR's right to explanation. They provide auditable, intuitive justifications for specific decisions. An auditor can verify that a model's counterfactual explanations are consistent, non-discriminatory, and based on permissible features.

• Demonstrate due diligence in automated decision-making.
• Create an audit trail of model reasoning for specific cases.
• Satisfy 'meaningful information' requirements under law.

By systematically generating counterfactuals for protected groups, analysts can probe a model for disparate treatment. For example, creating counterfactual profiles that differ only by gender can reveal if the model's output changes unjustly. This analysis is strengthened when anchored to a knowledge graph that defines protected attributes and fair relationships.

• Isolate the impact of sensitive attributes.
• Quantify fairness metrics using counterfactual distributions.
• Test for causal discrimination beyond statistical correlation.

Counterfactuals make AI decisions more comprehensible and persuasive to end-users, fostering trust and enabling effective collaboration. A doctor is more likely to trust an AI diagnostic suggestion if the system can explain: "The model suggests Condition A. If the patient's symptom Y were absent, the prediction would shift to Condition B."

• Bridge the cognitive gap between complex models and human intuition.
• Facilitate informed oversight and override by human experts.
• Use contrastive framing ("why this, not that?") for intuitive understanding.

In neuro-symbolic or graph-based RAG systems, the knowledge graph itself provides the constraints and ontology for generating realistic and feasible counterfactuals. The system ensures suggested changes respect ontological rules (e.g., a person cannot change their birth date) and entity relationships.

• Ground explanations in verified enterprise facts.
• Ensure feasibility by traversing valid graph relationships.
• Generate explanations that are consistent with the organization's semantic model.

Algorithmic Recourse provides actionable recommendations to an individual on how to change their input features to receive a more favorable outcome from an automated decision system. It is the practical application of a counterfactual explanation.

• Key Difference: While a counterfactual states "What minimal change would flip the prediction?", recourse provides the actionable steps to achieve that change (e.g., "Increase your credit score by 30 points").
• Use Case: In credit scoring, a counterfactual explanation identifies that a higher income would lead to loan approval; algorithmic recourse provides the specific, feasible path to increase income.

A Contrastive Explanation answers a 'why P rather than Q?' question by highlighting the features that differentiate the actual outcome from a plausible alternative outcome. It makes explanations more intuitive by focusing on a relevant comparison.

• Mechanism: Instead of explaining the prediction in absolute terms, it explains it relative to a chosen foil. For example, "Why was the loan rejected rather than approved? Because the debt-to-income ratio is 45%, whereas approved applications typically have a ratio below 35%."
• Relation to Counterfactuals: A counterfactual is a specific type of contrastive explanation where the foil is the desired outcome (e.g., approval).

A Causal Explanation identifies the cause-and-effect relationships within the input data that led to a model's prediction. It moves beyond correlation to establish a mechanistic, often validated, understanding of influence.

• Foundation: Often derived from or validated against a Structural Causal Model (SCM) or a causal knowledge graph, which encodes known cause-effect relationships.
• vs. Counterfactuals: A counterfactual explanation ("What if X had been different?") is inherently a causal question. High-quality counterfactuals should be causally plausible, not just statistically likely, requiring causal graphs to generate valid interventions.

This distinction categorizes explanations based on their scope. A local explanation justifies a single prediction for a specific instance, while a global explanation describes the overall behavior or logic of a model across its entire input space.

• Counterfactuals are Local: They are generated for a single data point (e.g., one customer's loan application).
• Global Analogue: A global, model-level understanding might involve generating multiple diverse counterfactuals across the dataset to infer general model behavior, or using a surrogate model to approximate global logic.

The Faithfulness Metric is a quantitative evaluation that measures how accurately a post-hoc explanation reflects the true decision-making process of the underlying black-box model. It assesses whether the explanation is a reliable proxy.

• Measurement: For a counterfactual explanation, faithfulness can be tested by perturbing the input according to the explanation and observing if the model's prediction changes as expected. A high correlation between the explanation's suggested changes and the model's actual sensitivity to those changes indicates high faithfulness.
• Importance: An unfaithful counterfactual is misleading and provides no real insight into the model's mechanics.

A Surrogate Model is a simple, interpretable model (e.g., a decision tree, linear model) trained to approximate the predictions of a complex, black-box model. It acts as a transparent stand-in for generating explanations.

• Role in Explainability: Explanations (including counterfactuals) are generated from the surrogate model, under the assumption it faithfully mimics the original model's behavior locally or globally.
• Connection: Techniques like LIME construct a local surrogate model (e.g., a linear model) around a specific prediction. A counterfactual for the black-box model can be derived by finding the minimal change that flips the surrogate model's prediction.