Global Versus Local Investigations

Opening Scene

A detective bureau runs two kinds of investigations at once: broad pattern analysis across every case of a certain type, and deep, specific work on one individual case. Both matter, and neither substitutes for the other. Pattern analysis reveals systemic issues a single case can’t show; a single case’s specifics reveal things a broad pattern summary washes out entirely.

Explaining a model works the same way, across two genuinely different scopes.

In Plain English

Global interpretability describes how a model behaves overall — which features matter most on average, what general patterns it has learned, whether it treats different subgroups consistently. Local interpretability describes why the model made one specific decision for one specific input. Feature importance from Article 4 is global; SHAP and LIME from Articles 5 and 6 are local. A real investigation, this series has argued from the start, needs both.

The Old Way

Before this distinction had formal names, the same global-versus-local split showed up wherever people studied any complex system:

A public health researcher studying disease patterns across a population, versus a doctor diagnosing one specific patient — both valid, both answering different questions.
A city planner studying traffic patterns across an entire road network, versus investigating one specific accident.
A quality control team analyzing a factory’s overall defect rate, versus investigating one specific defective unit.

In each field, the population-level and individual-level investigations were understood as complementary long before machine learning needed the same distinction.

What’s Changing (and Why AI Is the Reason)

As explainability techniques have matured, global and local methods have become distinct, well-developed toolkits rather than a single blurred category, letting investigators choose the right scope deliberately rather than defaulting to whichever tool happens to be available.
Global methods increasingly serve as an early-warning system for systemic issues — like the proxy-variable problems raised in Article 4 — while local methods handle individual accountability, connecting directly to the counterfactual explanations in Article 7.
Tooling now often surfaces both views side by side in the same dashboard, making it easier for a team to move fluidly between “how does this model behave overall” and “why did it decide this one case” without switching tools entirely.

The Metaphor, Fully Extended

The Investigation	Global vs. Local Concept
Pattern analysis across every case of a type	Global interpretability — overall model behavior
Deep investigation into one specific case	Local interpretability — one specific decision
A systemic issue only visible across many cases	A pattern only feature importance can reveal
A specific injustice only visible in one case’s details	A specific misclassification only SHAP or LIME can reveal

For Beginners: What to Actually Do

Learn to ask explicitly, for any interpretability question, “is this a global or a local question?” before reaching for a technique.
Practice moving between both scopes on the same model — check overall feature importance, then drill into one specific prediction.
Recognize that a model can look reasonable globally while still making troubling individual decisions, and vice versa.

For Practitioners and Leaders: The Deeper Layer

Build monitoring that tracks both global and local interpretability signals — a shift in either can signal a genuine problem the other view would miss.
Use global methods to catch systemic bias or drift early, and local methods to handle individual accountability and appeals.
Train stakeholders to ask the right scope question before requesting an explanation — much of this series’ communication challenge, covered in Article 16, starts with getting this distinction right.

Quick Recap

Global interpretability explains overall model behavior; local interpretability explains one specific decision.
Both scopes are genuinely necessary and complementary, not competing choices.
Global methods catch systemic issues; local methods handle individual accountability.
A model can appear reasonable at one scope while showing real problems at the other.

Where This Fits in the Series

This article named the global-versus-local distinction running through the whole series. Article 9 introduces a technique that sits deliberately between the two.