What is the difference between global and local SHAP explanations?

Global SHAP explanations describe how a model behaves across many predictions, so they are useful for spotting broad patterns and dominant features. Local SHAP explanations focus on one specific prediction and show which inputs pushed that result up or down. Practitioners use global views for model oversight and local views for case-by-case review.

How Global SHAP and Local SHAP Answer Different Questions

Global SHAP is the macro view: it aggregates SHAP values across many rows so you can see which features usually matter most, whether a feature tends to push predictions up or down, and whether the model behaves consistently across the dataset. Local SHAP is the micro view: it explains one prediction at a time, showing the specific feature contributions behind that single output.

That difference matters because the two views are not interchangeable. Global explanations help you understand model structure, feature dominance, and possible reliance on unstable signals. Local explanations help you review a single decision, challenge an unexpected outcome, and trace the features that drove one case rather than the model overall.

In practice, the same feature can look benign globally but still dominate a risky individual prediction, or look important globally while being irrelevant for a given case. That is why practitioners should treat global SHAP as an oversight tool and local SHAP as a case-review tool, not as competing versions of the same explanation.

When Each View Is the Right Tool

Use global SHAP when you need to compare patterns across a model, especially during validation, drift checks, feature review, or model governance. It is the right lens when the question is “What is this model generally doing?” rather than “Why did this one record receive this score?” Global views are also better for spotting dominant drivers that may deserve fairness review, data-quality investigation, or feature reduction.

Use local SHAP when the decision itself is under scrutiny, such as a high-impact approval, a false positive, a borderline classification, or a user-facing explanation. Local values are most useful when you need a defensible, row-level reason that can be inspected by an analyst, auditor, or reviewer. They are strongest when paired with the exact input values for the prediction being explained.

For model oversight, the key judgement is whether the explanation answer needs to generalize or to justify one specific output. If you are trying to manage the model, start global. If you are trying to review an outcome, start local. Many teams need both because a model can be globally reasonable while still producing problematic individual decisions.

Risk and Threat Considerations

SHAP is only as trustworthy as the model, feature set, and data distribution behind it. A global summary can hide sparse but severe failure modes, while a local explanation can overstate confidence in one prediction if the input data is unusual, incomplete, or already shifted from training conditions.

Failure mechanism: Global views can mask minority-pattern errors, and local views can be misread as causal proof when they are only attribution scores. If feature leakage, correlated inputs, or drift are present, the explanation may look coherent while the model is still brittle.

Impact: Teams may approve a model that appears stable in aggregate but fails on edge cases, or they may over-trust an isolated explanation and miss a broader modelling problem. In regulated or high-impact settings, that can translate into weak oversight, poor challengeability, and avoidable decision errors.

Standards & Framework Alignment

This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.

NIST CSF 2.0 and NIST AI RMF set the technical controls, while ISO/IEC 42001:2023 define the regulatory obligations.

Framework	Control / Reference	Relevance
NIST CSF 2.0	GV.RM-01 — Risk Management Strategy	SHAP supports model oversight and risk-based review of decision systems.
Recommendation — Use SHAP outputs to inform risk-based oversight of model behaviour and decision impact.
NIST AI RMF	GOVERN — Govern	Explaining model behaviour and individual outputs is part of AI governance and accountability.
	MAP — Map	Global and local explanations help map intended use, features, and impacts of the model.
	MEASURE — Measure	SHAP is used to measure feature influence and consistency across model behaviour.
Recommendation — Govern model explanation practices so global and local outputs support accountable AI oversight. Map the model’s features and use context before relying on SHAP explanations for decisions. Measure explanation stability across slices and time to detect brittle or shifting model behaviour.
ISO/IEC 42001:2023	A.5 — AI system impact assessment	SHAP evidence supports impact assessment of how AI systems influence outcomes.
Recommendation — Document SHAP-based impact assessments to support AI governance reviews and approvals.

Practitioner Guidance

What to verify: Check that the global ranking is stable across slices, time periods, and material subpopulations before treating it as model truth. Then confirm that the local explanation matches the exact feature values and preprocessing used for the scored record, not a simplified copy of the input.

Decision rule: If the question is about model behaviour, feature dominance, or oversight, use global SHAP first; if the question is about one outcome, use local SHAP first. When a local explanation conflicts with the global pattern, investigate data quality, correlated features, or drift before assuming the model is behaving correctly.

Practitioner takeaway: Global SHAP tells you whether the model is broadly explainable, but local SHAP tells you whether one decision is defensible; mature teams use the global view to govern the model and the local view to challenge individual outputs.