Completeness Score

A completeness score is a quantitative metric that evaluates the degree to which a feature attribution or explanation accounts for the total prediction made by a model. Its core purpose is to answer: 'Does this explanation capture all the important reasons for this prediction?' It is calculated by comparing the sum of the importance scores assigned by the explanation to the actual model output deviation from a baseline. A low score indicates the explanation is missing key predictive factors.

The score is often grounded in the efficiency axiom from cooperative game theory, as applied by methods like SHAP. Formally, for a model f and explanation φ, completeness requires: f(x) - f(x') = Σ φ_i, where x is the input instance, x' is a baseline input (e.g., all features masked), and φ_i is the attribution for feature i. The completeness score C can be expressed as C = 1 - |(f(x) - f(x')) - Σ φ_i| / |f(x) - f(x')|. A perfect score of 1.0 indicates the attributions sum exactly to the model's output difference.

Completeness is one of three pillars of explanation quality, distinct from but related to faithfulness and sufficiency.

• Faithfulness: Measures if the explanation's ranked importance correlates with the actual impact of features on the model's output. An explanation can be faithful but incomplete.
• Sufficiency: Measures if the top-K features identified by an explanation are, by themselves, sufficient for the model to make the same prediction. Sufficiency is a related but different property.
• Completeness: Ensures the magnitude of all attributions accounts for the total model output. A complete explanation is necessary for full accountability.

A primary method for empirically validating completeness is systematic perturbation analysis. This involves:

• Occlusion Tests: Iteratively removing or masking features deemed important by the explanation and observing the prediction change.
• Expected Change Calculation: The sum of prediction changes from occluding each feature should approximate the total deviation of the prediction from its baseline.
• Gap Measurement: The discrepancy between the expected total change (from the explanation) and the actual total model output change is the infidelity, which is inversely related to completeness. A high infidelity score indicates low completeness.

Achieving a perfect completeness score in practice is challenging due to:

• Non-Linear Interactions: In complex models, feature effects are not additive. The impact of occluding two features together may not equal the sum of occluding them individually, violating the linear assumption of some attribution methods.
• Baseline Sensitivity: The score is sensitive to the choice of baseline input (e.g., all zeros, average values). Different baselines yield different output deviations (f(x)-f(x')), changing the completeness calculation.
• Explanation Method Constraints: Some popular explanation methods (e.g., LIME) are not designed to be inherently complete. Their local surrogate models may not perfectly recover the complex model's output, leading to an inherent completeness gap.

For regulated industries and high-stakes applications, completeness is a non-negotiable component of the explanation audit trail. It provides:

• Quantitative Justification: Offers a numerical score to prove an explanation is comprehensive, moving beyond qualitative assessment.
• Regulatory Compliance: Supports adherence to principles like 'right to explanation' in regulations such as the EU AI Act, by demonstrating that explanations are thorough and account for the full decision.
• Risk Mitigation: Incomplete explanations can hide model reliance on spurious correlations or protected attributes. A high completeness score increases confidence that all decision drivers are visible for human review and bias auditing.

A faithfulness score is a quantitative metric that measures how accurately an explanation reflects the true reasoning process or causal factors of the underlying model for a given prediction. It is the core property that completeness, sufficiency, and infidelity all aim to quantify.

• Direct Measurement: Often evaluated via perturbation analysis, where features deemed important by the explanation are modified to see if the prediction changes as expected.
• Contrast with Completeness: While completeness asks 'did the explanation include all important features?', faithfulness asks a broader question: 'is the explanation's entire importance ranking correct?'

Sufficiency is an explanation metric that measures whether the subset of features identified as most important by an explanation is, by itself, sufficient for the model to make its original prediction. It tests for predictive power of the highlighted features.

• Evaluation Method: The top-k most important features (according to the explanation) are fed to the model, often with other features masked or set to a baseline. A high sufficiency score means this subset alone yields a prediction nearly identical to the full-input prediction.
• Relation to Completeness: Completeness and sufficiency are complementary. A complete explanation that includes all important features will naturally be sufficient, but a sufficient explanation may not be complete if it identifies a powerful but incomplete subset.

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 importance scores. It is a measure of explanation error.

• Calculation: Given an importance score vector, significant perturbations (e.g., removing top features) are applied to the input. Infidelity measures the expected difference between (1) the dot product of the importance vector and the perturbation vector, and (2) the actual change in the model's output.
• Inverse Relationship: A low infidelity score indicates high faithfulness. It directly penalizes explanations where the attributed importance does not correlate with the actual impact on the model's prediction under perturbation.

Perturbation analysis is a foundational explanation validation technique that systematically modifies or removes input features to observe the resulting changes in the model's output. It is the empirical basis for calculating completeness, faithfulness, and sufficiency.

• Methods: Includes occlusion sensitivity (for images), feature ablation, or adding noise.
• Ground Truth Proxy: The change in prediction due to a perturbation is treated as a proxy for that feature's 'true' importance, against which explanation methods are benchmarked.
• Challenges: Requires careful design of meaningful perturbations and baselines to avoid misleading results.

Explanation robustness refers to the property of an explanation method to produce consistent and stable attributions for a given prediction when the input or model is subjected to minor, semantically-preserving perturbations. It is concerned with the reliability of the explanation method itself.

• Stability Score: A related metric that quantifies this consistency.
• Contrast with Faithfulness: An explanation can be robust (consistent under small noise) but not faithful (consistently wrong). Conversely, a faithful explanation should be robust to semantically-invariant changes.
• Importance: Lack of robustness undermines user trust, as similar inputs yield wildly different explanations.

Local fidelity is a property of a post-hoc explanation that measures how well the explanation approximates the behavior of the complex model in the immediate vicinity of a specific input instance. It is the foundational promise of local explanation methods like LIME.

• Definition: A high-fidelity explanation acts as a locally accurate surrogate model.
• Evaluation: Measured by how well the explanation's simplified model (e.g., a linear model) predicts the outputs of the black-box model for perturbed samples near the instance of interest.
• Foundation for Validation: Metrics like completeness are only meaningful if the explanation method has high local fidelity to begin with; otherwise, it is explaining a surrogate that doesn't match the original model.