What is the difference between global and local SHAP explanations?

How global SHAP explains the model as a whole

Global SHAP explanations answer a population-level question: which features generally matter most, and in what direction they tend to move predictions. They are useful when you need an overview of model behaviour, such as identifying dominant drivers, comparing feature influence across the dataset, or spotting whether the model is relying on a narrow set of inputs too heavily. A global view is about patterns, not individual cases.

Because the explanation aggregates many predictions, it helps you judge whether the model is behaving in a way that matches domain expectations. For example, a feature can be globally important even if it does not move every prediction in the same direction. That makes global SHAP more suitable for model review, feature prioritisation, and governance conversations than for explaining a single decision to an end user.

How local SHAP explains one prediction

Local SHAP explanations answer a case-specific question: why did this record receive this score, label, or output? They break one prediction into feature contributions, showing which inputs pushed the result up, which pushed it down, and how those effects combine relative to the model baseline. This is the view practitioners use when they need to justify one outcome, investigate an anomaly, or compare a model decision with the underlying facts of the case.

Local explanations are most helpful when the concern is not model behaviour in general but the specific path to one result. They can reveal that a feature with modest overall importance had a large effect in one instance because the input values were unusual. That makes local SHAP a better fit for case review, dispute handling, and debugging surprising outputs than for broad model oversight.

When to use global versus local SHAP

The practical difference is the level of decision you are trying to support. Use global SHAP when you want to understand model behaviour across many records, validate whether the model is learning sensible broad patterns, or compare the relative importance of features over the full dataset. Use local SHAP when you need to explain one output, trace the feature-level reasoning behind a specific prediction, or support a human review of an individual case.

These two views are complementary rather than competing. Global SHAP can show that the model usually depends on a small set of variables, while local SHAP can show that a particular prediction was driven by an exception, interaction, or outlier value. Practitioners often move between the two: first confirm the model is broadly sensible, then inspect the individual prediction when a result looks unexpected.

Practitioner Guidance

What to verify: Check whether the global ranking of feature importance matches the business logic you expect before relying on local explanations for review. If the global story looks unstable, local attributions may still be mathematically valid but less trustworthy as evidence of model quality.

Decision rule: If you are answering “how does this model usually work?”, use global SHAP; if you are answering “why did this case get this result?”, use local SHAP. Do not treat one as a substitute for the other, because they answer different questions at different levels of granularity.

Practitioner takeaway: Global SHAP is for model oversight, local SHAP is for case explanation, and the strongest reviews use both, first to test the model’s general behaviour and then to inspect individual decisions when needed.

Related resources from NHI Mgmt Group

• How should teams use SHAP values when they need both global and local explanations of a machine learning model?
• What is the difference between global identity strategy and local governance?
• Why can SHAP explanations become misleading in real datasets?
• How should security teams evaluate whether a data lineage solution is truly global rather than just endpoint local?