A local explanation accounts for a single output

What it is

A holiday photo is tagged “beach.” A local explanation speaks only to this one: sand-coloured regions, a horizon, no indoor furniture, so the label lands on beach. It does not promise that the next night shot, or the next pool, will use the same weights. A local explanation accounts for a single output: the scope is locked to the instance in front of you, and does not generalise into the system’s character.

Ribeiro et al.’s LIME is the typical method at this layer — fit a readable surrogate near this instance, rather than draw the whole model.

Why it happens

A complex model bends in input space; the same feature can push opposite ways in different neighbourhoods. Local methods sample, weight, and fit only near the current point, and what they get is a temporarily tenable account of that neighbourhood. The “because of sand” the user sees is a nearby account of this classification, not “this system calls beach whenever it sees sand.”

Reading a local account as a product manual is a scope error. Once the scope is wrong, people use this instance to predict the next, and then fall on the next. Local methods exist precisely because the global surface is too wrinkled for one story to walk the whole way.

Studying it

Fix one output, supply a local surrogate (a LIME-style neighbourhood fit) or not, then give new samples from the same model and see whether people can predict the new label and point to the regions that mattered in this instance. Independent variables: local account present or not, neighbourhood width, distance of the new sample from this instance. Dependent variables: identification of factors in this instance, calibration of predictions on new samples, times the rule from this instance is misapplied to far samples.

The neighbourhood must actually be local. If sampling spreads far, a nominally local explanation is already leaking global structure and the task is contaminated.

Where it stops holding

If the model is close to globally linear, local and global coincide and this scope warning weakens. If the user’s task is only to handle the item in front of them (pass or bounce this image), local is the right grain. If the task is whether to ship the classifier, local is not enough. This entry does not turn a local explanation into an appeal procedure, and it does not discuss whether feature importance is causal.

Applying it

• Title the scope: “about this image,” not “how this system recognises scenes.”
• Put a limit next to local factors: on another image these factors may not hold.
• Do not turn this instance’s factor list into a reusable checklist unless you have checked that it stays stable outside the neighbourhood.
• Check: after the local account, show an image that looks similar but has a different label. If they insist on explaining the new image with the factors they just saw, and with high confidence, the scope has been read as global.

Related

• Same group: L5.02.2 A global explanation accounts for overall behavioural tendency · L5.02.3 Users need different explanation levels in different situations · L5.02.4 Local explanations support accepting or contesting a single result · L5.02.5 Global explanations support judging whether to keep relying on the system · L5.02.6 An overall impression induced from a few local explanations is often wrong · L5.02.7 Feature importance is a correlation ranking, not a causal account · L5.02.8 Developers need debugging information; end users need grounds for action — they cannot share one explanation
• Nearby: L5.01 Types of Explainability · L5.08 Counterfactual Explanations · L5.05 The Moderation Principle of Transparency
• Search terms: local explanation · LIME · instance-level explanation