Local vs. Global Explanations

The fundamental distinction lies in the scope of the analysis.

• Local Explanation: Focuses on a single data instance (e.g., one loan application, one molecule, one graph node). It answers: "Why did the model make this specific prediction for this specific input?"
• Global Explanation: Characterizes the model's overall logic or average behavior across the entire dataset or population. It answers: "What general patterns or rules does the model follow?" or "Which features are most important on average?"

The choice between local and global explanations is driven by the stakeholder's goal.

Local explanations are critical for:

• Debugging individual predictions (e.g., Why was this transaction flagged as fraud?)
• Providing actionable feedback to an end-user (Algorithmic Recourse).
• Validating a single high-stakes decision in regulated domains (e.g., healthcare, finance).

Global explanations are essential for:

• Model validation and trust during development and auditing.
• Understanding learned representations and feature relationships.
• Ensuring the model aligns with domain knowledge and business logic.

Different technical methods are optimized for local or global analysis.

Local Explanation Methods:

• LIME: Creates a local, interpretable surrogate model.
• SHAP (local Shapley values): Attributes the prediction for an instance to each feature.
• Counterfactual Explanations: Find minimal changes to alter the prediction.
• Saliency/Attention Maps: Highlight important input regions (pixels, nodes, tokens).

Global Explanation Methods:

• Global feature importance (e.g., mean absolute SHAP values).
• Partial Dependence Plots (PDPs).
• Rule extraction (e.g., mining decision rules from the model).
• Concept Activation Vectors (CAVs) for understanding learned concepts.

Knowledge Graphs (KGs) provide a structured semantic layer that enhances both local and global explanations.

• For Local Explanations: A KG can ground the features of a specific instance (e.g., a 'patient' node) in a rich network of related entities (medications, conditions, procedures). The explanation can trace the prediction through this semantic subgraph, making it more intuitive.
• For Global Explanations: The KG's ontology defines the concepts and relationships the model should respect. Global explanations can be validated against this ontology, and rule-based explanations can be expressed in the KG's formal language (e.g., OWL, SPARQL). This is a core tenet of Neuro-Symbolic AI.

The quality of explanations is measured differently based on scope.

Local Explanation Metrics:

• Faithfulness/Fidelity: Does removing features deemed important by the explanation significantly change the local prediction?
• Stability: Does the explanation remain consistent for similar inputs?
• Actionability: Can the explanation guide a user to a desired outcome (recourse)?

Global Explanation Metrics:

• Comprehensiveness: Does the explanation capture the model's major decision pathways?
• Consistency with Domain Knowledge: Does the extracted logic align with expert rules or the KG ontology?
• Representativeness: Is the explanation valid across diverse data subsets?

Local and global explanations are not mutually exclusive; they offer complementary views.

Key Trade-offs:

• Local explanations can be highly accurate for one case but may not reveal systemic model biases.
• Global explanations provide the big picture but can oversimplify and miss local nuances or edge cases.

A robust XAI strategy requires both. For example:

1. Use a global explanation to verify the model generally prioritizes clinically relevant features from a medical ontology.
2. Use a local explanation to audit why it recommended a specific treatment for Patient X, tracing the reasoning through their connected health data in the KG. This dual approach satisfies both model auditing (global) and individual right to explanation (local) requirements.

Explainable AI (XAI) is the overarching field of artificial intelligence focused on developing techniques to make the outputs and internal logic of machine learning models understandable to human users. It encompasses both local (single-instance) and global (whole-model) explanation methods. The goal is to build trust, facilitate debugging, and ensure compliance with regulations like the EU AI Act.

A Post-hoc Explanation is generated after a model makes a prediction, using a separate analytical method to interpret the black-box model's output. Most local explanation techniques (e.g., LIME, SHAP) are post-hoc. They are distinct from intrinsic explainability, where the model itself is inherently interpretable (e.g., a decision tree).

A Model-Agnostic Explanation method can generate interpretations for any machine learning model without requiring internal access to its architecture or parameters. Key examples include:

• LIME: Fits a simple local surrogate model.
• SHAP: Uses game-theoretic Shapley values. These methods are essential for explaining complex models like deep neural networks and are applicable to both local and global scopes.

A Surrogate Model is a simple, interpretable model (e.g., a linear regression, decision tree, or rule set) trained to approximate the predictions of a complex, opaque model. It is a core technique for generating explanations:

• Locally: A surrogate is trained on perturbed samples around a single instance.
• Globally: A surrogate is trained to mimic the complex model's behavior across the entire dataset, providing a comprehensible overview of its logic.

The Faithfulness Metric is a critical evaluation measure for post-hoc explanations. It quantifies how accurately the explanation reflects the actual reasoning of the underlying model. For a local explanation, faithfulness is tested by perturbing the features deemed important and measuring the correlation between the perturbation's impact on the prediction and the explanation's importance scores. A high-fidelity explanation is one where the highlighted features are truly causal to the model's output.

These are distinct but related concepts:

• Interpretability refers to the extent to which a human can understand the cause of a model's decision by examining its internal mechanics (e.g., weights in a linear model). It is an intrinsic property.
• Explainability refers to the ability to provide understandable reasons for a model's behavior, often using external, post-hoc methods. A global explanation aims for interpretability of the model's overall logic, while a local explanation provides explainability for a specific case.