Contrastive Explanations

A contrastive explanation is inherently counterfactual. It does not simply list features important for the prediction; it identifies the minimal changes required to flip the prediction from the observed outcome P to the foil Q. For example, explaining 'Why was the loan denied rather than approved?' might highlight that increasing the applicant's income by $10,000 would have changed the outcome. This focuses on actionable, decision-relevant factors.

The utility of a contrastive explanation is entirely dependent on the choice of the foil (the 'Q' in 'why P rather than Q?'). A meaningful foil is a plausible alternative outcome.

• Good Foil: 'Why was this image classified as a wolf rather than a husky?' (plausible confusion).
• Poor Foil: 'Why was this image classified as a wolf rather than a toaster?' (implausible). The explanation highlights features that discriminate between the specific classes of P and Q, such as background context or snout shape, not all features relevant to 'wolf-ness'.

Contrastive explanations are typically sparse, identifying only the few features that are selectively relevant to the contrast. They filter out features that are equally present or absent in both P and Q scenarios. If a credit model uses 100 features, a contrastive explanation for 'denied vs. approved' might isolate only 3-5 features where the applicant's profile meaningfully differed from the approval threshold. This aligns with human cognitive biases for simple, selective causes.

By framing the explanation around a change in outcome, contrastive explanations suggest causal, actionable insights. They answer the user's implicit question: 'What could I change to get a different result?' This makes them particularly valuable for:

• Recourse: Telling a user what to modify for a favorable decision.
• Debugging: Helping a developer understand what specific feature interactions cause a model error.
• Regulatory Compliance: Providing reasons that directly relate to adverse decisions under laws like GDPR.

The faithfulness of a contrastive explanation is directly testable using perturbation analysis. The method suggests that changing features F will flip the prediction from P to Q. Validation involves:

1. Creating a perturbed input where features F are modified as suggested.
2. Feeding it to the model.
3. Verifying the output becomes Q. A high sufficiency score confirms the explanation's causal claim. This empirical test is a core component of explanation score validation.

Contrastive explanations complement but differ from other explainability techniques:

• vs. SHAP/LIME (Feature Attribution): These assign importance scores for a single prediction P. Contrastive explanations require a foil Q and highlight discriminative importance between two outcomes.
• vs. Counterfactual Explanations: These are a subset of contrastive explanations. A counterfactual is a contrastive explanation where the foil Q is a desired outcome (e.g., 'What changes would get me approved?'). All counterfactuals are contrastive, but not all contrastive explanations are counterfactuals (e.g., 'Why wolf vs. husky?' doesn't imply a desired change).

Scenario: A model denies a loan application (Prediction P: 'Deny'). The applicant asks, 'Why was I denied, rather than approved?'

Contrastive Explanation: 'Your application was denied rather than approved because your debt-to-income ratio is 45%, which exceeds our approval threshold of 35%. If your ratio were below 35%, your application would likely have been approved, even with your current credit score.'

• Key Feature: Debt-to-income ratio.
• Contrast Case (Q): A hypothetical scenario where the ratio is ≤35%.
• Mechanism: The explanation isolates the single most decisive feature that flips the prediction from the desired outcome to the actual one.

Scenario: A convolutional neural network classifies a skin lesion image as malignant melanoma (P) rather than benign nevus (Q).

Contrastive Explanation: 'The lesion is classified as melanoma rather than a benign mole primarily due to the highly irregular border and the presence of multiple colors within the lesion. A benign nevus typically exhibits a smooth, regular border and a more uniform pigmentation.'

• Key Features: Border irregularity and color variegation.
• Contrast Case (Q): The prototypical features of a benign nevus.
• Utility: Directly addresses a clinician's counterfactual question, focusing on discriminative visual features that align with medical expertise.

Scenario: An e-commerce platform's model recommends a high-end DSLR camera (P) to a user instead of a smartphone (Q), which was the user's expected recommendation.

Contrastive Explanation: 'We recommended the DSLR rather than a smartphone because your browsing history shows repeated visits to professional photography tutorials and reviews for interchangeable-lens cameras. A smartphone recommendation is typically driven by searches for 'portable photography' or 'social media,' which are absent from your recent activity.'

• Key Features: Browsing history semantic content.
• Contrast Case (Q): The user profile that typically triggers a smartphone recommendation.
• Actionability: Explains the system's reasoning by contrasting the user's actual signal against the expected signal for the alternative outcome.

Scenario: A self-driving car's planning module decides to brake abruptly (P) instead of maintaining speed (Q) when a ball rolls into the street.

Contrastive Explanation: 'The vehicle initiated hard braking rather than continuing because the object was classified as a 'ball' with high confidence (92%). The system's policy associates balls with a high probability (>80%) of a child following. Maintaining speed is the policy output only when object classification confidence for 'debris' is above 95%.'

• Key Features: Object classification (ball) and associated risk probability.
• Contrast Case (Q): The scenario where the object is classified as low-risk debris.
• Causality: Highlights the specific perceptual classification and the downstream policy rule that creates the fork between the two possible actions.

Scenario: A moderation AI flags a social media post as 'hate speech' (P) instead of 'strong criticism' (Q).

Contrastive Explanation: 'This post was flagged as hate speech rather than strong criticism because it contains a dehumanizing metaphor targeting a protected group. Our model is trained to distinguish criticism, which focuses on actions or ideas, from hate speech, which attacks inherent attributes. Removing the dehumanizing metaphor while keeping the critical core would likely result in a 'strong criticism' classification.'

• Key Feature: Use of dehumanizing language.
• Contrast Case (Q): A minimally edited version of the post focusing on actions/ideas.
• Fairness & Appeal: Provides a clear, actionable path for the user to understand the boundary and modify content appropriately.

Scenario: A translation model renders the French phrase 'Je suis plein' into English as 'I am full' (P - from eating) instead of the intended 'I am pregnant' (Q - colloquial French).

Contrastive Explanation: 'The model translated this as 'I am full' rather than 'I am pregnant' because the immediate textual context provided no surrounding words related to pregnancy or motherhood. The model's most frequent training association for 'Je suis plein' in isolation is the literal 'I am full.' To get the pregnancy meaning, the context would need supporting terms like 'bébé' or 'attendre.'

• Key Feature: Absence of contextual semantic cues.
• Contrast Case (Q): The required contextual signals for the idiomatic interpretation.
• Debugging: Helps developers understand if the error stems from a contextual deficiency or a training data gap.

Counterfactual explanations answer 'What if?' by identifying the minimal changes to input features required to flip a model's prediction to a desired alternative class. Unlike contrastive explanations, which explain 'why P rather than Q', counterfactuals focus on actionable paths to a different outcome.

• Core Mechanism: Searches the input space for the nearest instance that results in a different prediction.
• Example: For a loan denial, a counterfactual might state: 'Your application would have been approved if your annual income were $5,000 higher.'
• Key Distinction: Provides a recipe for change, whereas a contrastive explanation justifies the existing prediction relative to a foil.

SHAP is a unified framework for feature attribution based on cooperative game theory. It assigns each feature an importance value (Shapley value) for a specific prediction, representing its average marginal contribution across all possible feature combinations.

• Theoretical Basis: Grounded in Shapley values from game theory, ensuring properties like local accuracy and consistency.
• Output: Produces a single number per feature, showing how much it pushed the prediction higher or lower from a baseline expectation.
• Relation to Contrastive: SHAP values can be used to construct contrastive explanations by comparing the Shapley value contributions for features that differ between the actual instance and a chosen contrast case.

The faithfulness score is a quantitative metric that measures how accurately a feature attribution or explanation reflects the true causal factors of the underlying model for a given prediction. It validates whether the explanation's highlighted features are genuinely important to the model.

• Common Measurement Method: Perturbation analysis. Features deemed important by the explanation are removed or altered, and the resulting drop in model prediction confidence is measured. A larger drop indicates higher faithfulness.
• Critical for Validation: A core metric in Explainability Score Validation, used to audit post-hoc explanations like contrastive or SHAP attributions.
• Direct Application: Evaluating a contrastive explanation involves checking if perturbing the 'discriminative features' (those that favored P over Q) causes the model's preference to diminish or reverse.

Perturbation analysis is a foundational technique for both generating and validating explanations. It involves systematically modifying input features and observing the resulting changes in the model's output to infer feature importance or test explanation robustness.

• Two Primary Uses:
  • Explanation Generation: Methods like LIME create explanations by perturbing inputs around an instance and fitting a simple local model.
  • Explanation Validation: Used to compute faithfulness scores and infidelity metrics by testing if perturbing important features causes significant prediction change.
• In Contrastive Context: To validate a contrastive explanation, one perturbs the features identified as key differentiators to see if the model's preference for P over Q disappears.

Anchors is a model-agnostic explanation method that provides a high-precision rule (an 'anchor')—a set of if-then conditions on input features—that 'anchors' the prediction, making it locally robust to changes in all other features.

• Explanation Format: 'IF [feature A = value X] AND [feature B > value Y], THEN the prediction is Z with high probability.'
• Contrastive Relationship: An anchor rule can be seen as a highly local and precise form of contrastive explanation. It defines a sufficient region in feature space for the prediction, implicitly contrasting against instances outside that rule-defined region.
• Key Property: Provides precision guarantees, indicating the probability that the prediction holds when the anchor conditions are met but other features are randomly perturbed.

Infidelity is an explanation metric that quantifies the degree to which an explanation fails to accurately reflect the model's output when the input is perturbed according to the explanation's own importance scores. It is a formal measure of unfaithfulness.

• Mathematical Definition: Measures the expected squared error between the explanation's attribution-based prediction and the actual model output change under meaningful perturbations.
• Interpretation: A low infidelity score indicates the explanation reliably predicts how the model will behave when inputs are changed, aligning with the explanation's importance weights.
• Validation Role: A core quantitative tool in post-hoc explanation validation. For a contrastive explanation, infidelity would assess how well the highlighted feature differences predict the model's comparative score between instance P and foil Q.