GFGE: Unifying Explainable AI Methods through an Interpretation Framework

1 Introduction

• Existing explanation methods draw on different evidence and answer different questions, while inconsistent interpretability and explainability concepts complicate coherent comparison.Users also seek to understand decisions, assess fairness, and identify possible changes to outcomes.Source: §1 · pp. 1–2
• The paper separates interpretation as making sense of information from explanation as communicating a selected account, and identifies why, why-not, what-if, how-to, and global-pattern questions.These are among the explanatory questions addressed; the list is not stated to be exhaustive.Source: §1 · p. 2
• The Interpret/Explain Schema connects interpretations of data, models, and outputs across developers, systems, and end-users.It describes information flow through the AI pipeline and makes the conceptual distinction computationally explicit.Source: §1 · p. 2
• GFGE operationalises the schema through five interchangeable roles and is compatible with intrinsic and post-hoc approaches without being tied to a model family or post-hoc technique.The roles are data, model, and output interpretation, optional post-hoc interpretation, and aggregation.Source: §1 · pp. 2–3
• GFGE is instantiated across attribution, counterfactual, surrogate, concept and prototype, argumentation, and language-model methods, including hybrid procedures.The comparison tests descriptive expressiveness rather than the empirical quality of generated explanations.Source: §1 · p. 3

2 Literature Review

• The review selects work on definitions, cross-method frameworks, and methods that expose boundary cases for GFGE, rather than aiming for an exhaustive survey.Its cases were chosen for conceptual diversity, so the literature set cannot support prevalence claims or comprehensive XAI coverage.Source: §2 · p. 3

2.1 A General View

• XAI has grown amid expectations that AI decisions be intelligible to stakeholders across increasingly broad application domains.The section frames transparency, accountability, and trust as concerns surrounding algorithmic decision-making.Source: §2.1 · p. 3
• Interpretability and explainability are used inconsistently across the literature, complicating the formulation, comparison, and evaluation of explanation methods.Some works distinguish the terms, while others use them interchangeably or treat the distinction as secondary.Source: §2.1 · pp. 3–4

2.2 To Explain or to Interpret?

• One position treats interpretability as a prerequisite for explainability, distinguishing model understanding from explaining a particular result.In this view, transparent internal mechanisms allow models to explain their outputs by default.Source: §2.2 · p. 4
• A second position treats explainable models as interpretable by default but allows that interpretable models may not yield valid explanations.This view emphasizes justification and correctness, often with a causal interpretation.Source: §2.2 · p. 4
• A third position uses interpretability and explainability interchangeably, sometimes referring to transparency or the ability to explain individual predictions.Taken together, the positions lack uniform, systematic grounding for structuring methods across models and contexts.Source: §2.2 · p. 4

2.3 Explanation: What?

• Explanations combine a process of constructing and selecting information with a product that addresses a question, such as why, what-if, how-to, or global-pattern.The question may also request a contrast or feasible change, and the account is shaped by recipient beliefs, expectations, and background knowledge.Source: §2.3 · p. 4
• Figure 1 illustrates an analyst reading evidence, giving an account, and a recipient interpreting it through a different framework.This distinction provides a conceptual basis for the IES.Source: Figure 1 · p. 5, §2.3 · p. 5
• Sense-reading interprets information through an individual mental model, while sense-giving assigns meaning to information conveyed to others.The recipient may interpret an account differently from the analyst’s intended meaning.Source: §2.3 · p. 5

2.4 Explanation: Why?

• Explanation drivers motivate frameworks, but do not resolve how explanations should be structured or stakeholders’ needs aligned.Source: §2.4 · p. 5

2.5 Explanation: How?

• Intrinsic versus post-hoc, local versus global, and model-specific versus model-agnostic distinctions help compare methods, while richer frameworks encode mechanisms, access, stakeholders, and interaction.Source: §2.5 · p. 5
• Hybrid methods can carry several taxonomic labels, motivating a process record of their ordered operations and assumptions rather than static categories alone.Source: §2.5 · pp. 5–6

• Evaluation strategies include application-grounded, human-grounded, and functionally-grounded approaches addressing usefulness, human understanding, and model fidelity.Source: §2.6 · p. 6
• Evaluation frameworks assess explanation quality, whereas GFGE focuses on the structure of explanation generation.Source: §2.6 · p. 6
• GFGE’s structural focus is compatible with existing evaluation methodologies.Source: §2.6 · p. 6

• Schwalbe and Finzel’s survey of surveys provides a detailed taxonomy of method properties, while Wang et al. connect stakeholder questions to methods across the AI lifecycle.Source: §2.7 · p. 6
• Other reviews support technique selection, distinguish interpretability mechanisms, organise human-centred design and evaluation, or examine LLM interpretability and interventions.Source: §2.7 · p. 6
• GFGE maps methods by their instantiated roles, transferred evidence, and transition from analyst sense-reading to recipient-oriented sense-giving.Source: §2.7 · p. 6
• GFGE complements existing taxonomies and user-centred evaluation rather than replacing them.Source: §2.7 · p. 6

• The reviewed literature includes method taxonomies, discipline-specific definitions of interpretability and explainability, and evaluation frameworks for explanation quality [21] [21] [31].Source: §2.8 · pp. 6–7
• The narrower gap is a common representation linking evidence provenance, the order of interpretative operations, and the transition from analyst findings to a recipient-facing account.Source: §2.8 · p. 7
• The proposed contributions combine working definitions grounded in sense-reading and sense-giving, an IES linking stakeholder interpretations across the AI pipeline, and a modular workflow for intrinsic, post-hoc, and hybrid methods.Source: §2.8 · p. 7
• This mapping complements method selection and evaluation, but does not establish that a mapped explanation is faithful or understood.Source: §2.8 · pp. 7–8

3 Interpretation and Explanation: Conceptual Basis

• Sense-reading is interpreting received information, while sense-giving assigns meaning to information conveyed to others.Different people may derive different meanings from the same display.Source: §3 · p. 8
• Interpretability concerns an observer deriving meaningful understanding; explainability concerns constructing and communicating an account intended to make an event or outcome understandable.The account may include reasons or causes when those claims are warranted.Source: §3 · p. 8
• An analyst may interpret data or model behaviour before explaining it, and a recipient interprets the account; communication does not guarantee successful understanding.An account of association is not automatically causal.Source: §3 · p. 8
• The IES locates interpretation and explanation in an AI system, while GFGE identifies operations that can connect them.Source: §3 · p. 8
• The definitions are conceptual and observer-relative, not measurement scales; whether an account becomes meaningful requires separate observation and evaluation.GFGE’s Interpret modules extract or organise candidate evidence, rather than performing a person’s cognitive interpretation.Source: §3 · p. 8

4.1 AI: To Interpret or to Explain?

• The IES distinguishes analyst sense-reading of data, model, and output from selecting an account for communication and the recipient’s interpretation of it.Recipient understanding is an empirical question, not an automatic consequence of exposing model information.Source: §4.1 · p. 9, Figure 2 · p. 9

4.2 Definitions in the Context of AI

• The five capacities distinguish data, model, result, post-hoc, and explanation interpretability, connecting analyst-available evidence to recipient sense-reading through sense-giving.These capacities do not guarantee fidelity, causal validity, or user benefit; those require separate evaluation [21].Source: §4.2 · p. 10

4.3 Discussion: Stakeholders and Explanatory Needs

• The same system can raise different questions for developers, domain experts, and affected users, so the relevant account depends on audience and task.Reasons for explaining include debugging, accountability, appropriate reliance, and decision support.Source: §4.3 · p. 10

4.4 General Framework for Generating Explanations

• GFGE operationalises the IES with data, model, output, optional post-hoc, and aggregation roles, while leaving their implementation open to different methods.The framework is not tied to a model family or post-hoc technique.Source: §4.4 · pp. 10, 11, 14, 15
• Evidence records link claims to their sources, producing operations, parent claims, scope, assumptions, and limitations, while retaining unresolved dependencies.Aggregation selects an account for a question and audience, and communicated claims retain links to their records.Source: §4.4 · pp. 11, 14, §4.5 · p. 16
• The operation graph records method-specific dependencies, which may place post-hoc analysis before model interpretation or repeat a role.Algorithm 1 shows a canonical single pass, not GFGE’s general semantics.Source: §4.4 · pp. 12, 14, Figure 3 · p. 13

4.5 Discussion

• GFGE is a structured narrative review’s common procedural representation, not a new graph formalism or a framework for defining explanation quality.Its scope includes intrinsic, post-hoc, and hybrid procedures, while fidelity and usefulness remain matters for independent evaluation [51].Source: §4.4 · p. 15, §4.5 · p. 16
• GFGE makes method composition and the analyst-to-recipient transition inspectable, representing hybrid workflows and identifying missing evidence relative to declared requirements.Structural fit does not establish correctness, fidelity, understanding, or usefulness.Source: §4.4 · p. 15, §4.5 · pp. 16–17

5.1 Post-Hoc Attribution Techniques: LIME and SHAP

• LIME uses local perturbations and a sparse surrogate, while SHAP records feature semantics and background populations and aggregates Shapley contributions.The section also maps attribution outputs to ranked lists, plots, or natural language.Source: §5.1 · pp. 17–18

5.2 Methods Built on Post-Hoc Attribution Models

• Attribution-guided counterfactual workflows combine post-hoc attribution with further reasoning; one recommender example prioritizes modifications to features with the strongest Shapley contributions.The resulting account can combine why an item was recommended with what feature changes would yield an alternative.Source: §5.2 · pp. 18–19

5.3 Model-Intrinsic Explanation Methods Based on Argumentation

• Argumentation-based recommenders encode explanations in the model, with support and attack relations interpreted as model evidence and translated into human-readable accounts.These methods do not require a separate post-hoc stage.Source: §5.3 · p. 19

5.4 Counterfactual and Recourse Methods

• Counterfactual workflows separate prediction-changing candidates from feasible actions by recording permitted changes and plausibility constraints, searching candidates, and reporting selected contrasts with assumptions.The framework distinguishes a changed prediction from an actionable intervention.Source: §5.4 · p. 20

5.5 Surrogate, Rule, Concept, and Prototype Methods

• A rule is post-hoc when it approximates another model but intrinsic when it drives prediction; concept and prototype methods likewise distinguish analytic constructions from learned mechanisms.Concept semantics and predictive fidelity remain separate evaluation claims.Source: §5.5 · p. 20

5.6 Language-Model and Recipient-Oriented Workflows

• Language models can render recorded claims into readable rationales, but new propositions require supporting evidence; audience also guides recipient-facing presentation.A readable rationale alone does not reveal the model’s actual decision process.Source: §5.6 · pp. 20–21

5.7 Discussion

• GFGE’s shared roles make method-specific evidence flows explicit without equating attributions, counterfactuals, rules, and generated rationales.The mappings are illustrative, not evidence of comparative superiority or completeness. The paper reports one author-coded pass over three selected cases, without annotator agreement statistics or a user study.Source: §5.7 · pp. 21–25

6 Conclusion

• GFGE operationalises the IES through data, model, and output interpretation, optional post-hoc analysis, and aggregation, guided by question, audience, and context.The IES connects an analyst’s interpretation of evidence with communication of an account and a recipient’s interpretation.Source: §6 · p. 25
• Role-typed operation graphs capture method-specific dependencies and repeated roles, while claim-level records connect communicated explanations to supporting operations and evidence.The framework preserves the distinct sources, assumptions, and scope of intrinsic, post-hoc, and hybrid methods.Source: §6 · p. 25
• GFGE supports comparison of explanation procedures and identification of unresolved dependencies relative to a specified question and declared evidential requirements.Future work should test reproducibility, method-specific evidence support, and recipient understanding and use.Source: §6 · p. 26