Category: Explainability & Interpretability

Global Interpretability

Also known as: global model interpretability, global interpretation

Simply put

Global interpretability refers to understanding how a machine learning model behaves overall, across its full range of inputs, rather than explaining a single prediction. It helps people see which features generally influence the model's outputs and how the model tends to behave on average across a dataset. This contrasts with local interpretability, which focuses on explaining an individual decision.

Formal definition

Global interpretability, as commonly defined, describes methods and analyses aimed at characterizing the average or aggregate behavior of a machine learning model across a dataset, including how features contribute to the model's outputs overall. Such global insight is typically used to support feature selection, model development, and general model understanding. It is distinguished from local interpretability, which addresses the explanation of individual predictions; recent work (for example, Bassan et al., 2024) has examined this local-versus-global distinction through computational complexity theory, indicating that the two perspectives can differ substantially in tractability. Note that interpretability more broadly is often framed as the degree to which a human can understand the cause of a decision, and definitions in this area remain varied across the literature; this entry does not address whether any particular global method is required or sufficient for regulatory or model risk management purposes.

Why it matters

Global interpretability matters because it addresses a different question than most individual explanation techniques: not why a single decision was made, but how a model tends to behave across its full range of inputs. Understanding which features generally influence a model's outputs, and how the model behaves on average across a dataset, supports model development activities such as feature selection and general model understanding. For teams responsible for understanding and documenting model behavior, this aggregate view can help characterize what a model has learned overall rather than reasoning only from case-by-case outputs.

The distinction between global and local interpretability is not merely conceptual. Recent work (for example, Bassan et al., 2024) has examined the local-versus-global distinction through computational complexity theory, indicating that the two perspectives can differ substantially in tractability. This suggests that a method or level of understanding available at one scope may not transfer cheaply to the other, and that practitioners should be explicit about which question they are answering. Treating a global summary of average behavior as if it explained any particular decision, or vice versa, risks misstating what the analysis actually supports.

Because interpretability is often framed as the degree to which a human can understand the cause of a decision, and because definitions in this area remain varied across the literature, global interpretability should be understood as one lens rather than a complete account of a model. This entry does not address whether any particular global method is required or sufficient for regulatory or model risk management purposes; those determinations depend on context, applicable frameworks, and the specific use of the model, and should not be inferred from the availability of a global interpretability technique alone.

Who it's relevant to

Data Scientists and Model Developers

Global interpretability directly supports development-stage activities such as feature selection and general model understanding by revealing how features contribute to outputs overall. Developers can use it to sanity-check whether a model's broad behavior aligns with domain expectations, while remaining careful to treat it as a description of average behavior rather than an explanation of individual predictions.

Model Risk Managers and Validators

Those tasked with understanding and documenting model behavior may find global interpretability useful for characterizing what a model has learned across a dataset. However, this entry does not establish whether any particular global method is required or sufficient for model risk management purposes; such judgments depend on the applicable framework and the specific use of the model, and the global-versus-local scope of any analysis should be stated explicitly.

Auditors and Reviewers

When assessing interpretability claims, auditors should confirm whether a stated explanation concerns aggregate model behavior (global) or a specific decision (local), since these answer different questions and can differ substantially in tractability. This helps prevent a global summary from being cited as evidence about an individual output, or vice versa.

Policy and Legal Specialists

Because interpretability is often framed as the degree to which a human can understand the cause of a decision, and definitions remain varied across the literature, policy and legal readers should be cautious about assuming that a global interpretability technique satisfies any particular requirement. The suitability of a method depends on context and applicable obligations, which fall outside the scope of this definition.

Inside Global Interpretability

Model-Wide Explanation Scope

Global interpretability concerns understanding how a model behaves across its entire input space or population, rather than explaining a single prediction. It seeks to characterize the general logic, dominant relationships, and overall decision structure of the model.

Feature Importance Aggregation

Techniques that summarize which features most influence model outputs across many observations, for example aggregated attribution measures. These indicate relative influence in aggregate but do not, on their own, establish causal relationships in the underlying data.

Surrogate or Approximation Models

Simpler, more transparent models sometimes fitted to approximate the behavior of a complex model at a global level. The fidelity of such approximations is itself a limitation, since the surrogate may not faithfully represent the original model everywhere.

Relationship and Response Characterization

Methods that describe how outputs typically respond to changes in inputs across the population, such as partial dependence style summaries. These reflect average or marginal behavior and can obscure heterogeneity across subgroups.

Distinction From Local Interpretability

Global interpretability addresses overall model behavior, whereas local interpretability explains individual predictions. The two are complementary; global summaries can mask instance-level behavior, and local explanations do not necessarily generalize.

Common questions

Answers to the questions practitioners most commonly ask about Global Interpretability.

Is global interpretability the same as explainability?

Not exactly. Global interpretability, as commonly defined, refers to understanding the overall behavior and decision logic of a model across its entire input space or population. Explainability is a broader term that includes both global and local approaches, and often refers to post-hoc techniques used to render otherwise opaque model behavior understandable. Professionals frequently blur these terms, but experts typically distinguish global interpretability (model-wide understanding) from local interpretability (explaining a single prediction) and from explainability as an umbrella concept. Treating them as interchangeable can obscure whether a given technique actually accounts for model-wide behavior or only individual outputs.

Does achieving global interpretability mean the model is inherently interpretable?

No. Global interpretability describes the goal of understanding a model's behavior across its input space, and this can be pursued for both inherently interpretable models (such as linear models or shallow decision trees) and complex models approximated through post-hoc methods. An inherently interpretable model is one whose structure is directly understandable, whereas global interpretability applied to a black-box model relies on approximations that may not fully reflect the underlying mechanics. Assuming that global interpretability equals inherent interpretability is a common error; a post-hoc global explanation can be incomplete or misleading if it diverges from the model's actual internal logic.

What techniques are commonly used to pursue global interpretability?

In many practices, approaches include using inherently interpretable model classes where feasible, and applying post-hoc methods such as global feature importance measures, partial dependence-style analyses, surrogate models trained to approximate a complex model's behavior, and aggregation of local explanations into a model-wide view. The appropriateness of each depends on the model type, use case, and the level of assurance required. Practitioners should note that surrogate and approximation methods carry fidelity limitations and may not accurately represent the target model in all regions of the input space.

How does global interpretability relate to model validation and model risk management?

Global interpretability can support model validation activities by helping validators assess whether a model's behavior is conceptually sound and consistent with domain expectations across the population, which is one input among many in a validation process. It can also inform ongoing monitoring within a model risk management framework. However, interpretability findings are not a substitute for other validation components such as outcome analysis, benchmarking, or testing of data and assumptions. Its role should be scoped as one line of evidence rather than as a standalone control.

How do global interpretability methods interact with local explanations in practice?

Local explanations address individual predictions, while global interpretability seeks model-wide understanding, so the two are typically used together rather than as alternatives. Some workflows aggregate many local explanations to approximate a global view, but this aggregation can mask heterogeneity in how a model behaves across different subpopulations or input regions. Practitioners often maintain both perspectives to avoid over-generalizing from a global summary or over-relying on isolated local cases.

What are the limitations to document when relying on global interpretability?

Documentation should note where global explanations are approximations rather than exact reflections of model logic, the regions of the input space where fidelity may be weaker, and the assumptions underlying any surrogate or aggregation method. It is also useful to state that global interpretability supports understanding and risk management but does not eliminate model risk, and that the adequacy of a given technique may vary by model type, sector, and use case. Where definitions or expectations are contested or evolving, that should be flagged explicitly rather than presented as settled.

Common misconceptions

Global interpretability is the same as interpretability in general, or interchangeable with explainability.

Global interpretability is one scope of understanding model behavior, distinct from local interpretability, which explains individual outputs. Interpretability, as commonly framed, refers to the degree to which a model's mechanics can be understood, while explainability often refers to post hoc techniques applied to explain outputs; practitioners frequently blur these, but they are not identical.

A global interpretability summary tells you why any specific prediction was made.

Global methods describe aggregate or typical behavior across the input space and can miss instance-level nuances, subgroup differences, and interactions. Individual decisions typically require local interpretability methods, and relying on global summaries alone can misrepresent specific cases.

Achieving global interpretability demonstrates that a model is fair, unbiased, or compliant.

Understanding overall model behavior is distinct from assessing bias, fairness, or regulatory conformance. Global feature importance indicates influence, not appropriateness or causation, and interpretability alone does not eliminate risk or satisfy governance and validation obligations.

Best practices

Pair global interpretability techniques with local, instance-level methods so that aggregate behavior and individual predictions are both examined rather than assuming one substitutes for the other.

State the scope and assumptions of each method used, distinguishing what reflects average or marginal behavior from what reflects behavior across specific subgroups or regions of the input space.

Assess and document the fidelity of any surrogate or approximation model, since a low-fidelity approximation can misrepresent the behavior of the original model.

Treat aggregated feature importance as a measure of influence within the model, not as evidence of causal relationships in the underlying data.

Examine heterogeneity across subgroups explicitly, because global summaries can mask differences in how the model responds for different populations.

Position interpretability findings as inputs to broader governance and validation activities rather than as standalone evidence of fairness, compliance, or reduced risk.