A global explanation accounts for overall behavioural tendency

What it is

A résumé screener that “overall down-weights candidates with career gaps” is not a sentence about one rejected letter. It is a bias the system shows across many examples. A global explanation accounts for overall behavioural tendency: it gives the apparatus’s temperament, not a nearby reason for this output.

It answers how this thing usually treats people. It does not answer why your file landed in this band. That question needs another scope.

Why it happens

Global accounts come from aggregating a distribution: partial dependence, rule lists, pass rates sliced by group, prototype cases. What they compress is an average effect. The average can be stable while a single point runs the other way — gaps cost most people points, and count as project experience for one returning researcher. If users read their own file through that temperament, they get the illusion that “the manual is right and my result is wrong,” which is a scope error.

Global methods exist because people have to decide whether to keep handing a class of tasks to this apparatus. That is a judgement of temperament, not of a single label. Neighbourhood surrogates such as LIME cannot give this temperament; force them and a few coincidentally similar points get mistaken for character.

Studying it

Give two groups a global tendency (group pass rates, main rules) versus several single-output local accounts, then have them estimate pass rates on a new group and the fate of one concrete file. Independent variables: scope of the account (global / local only), distance of the group from the training distribution. Dependent variables: calibration of the system’s tendency, over-extrapolation to a single file, willingness to keep using it on that group.

Global numbers must come from a slice in the same distribution as the user’s task. Passing off mean accuracy from a public benchmark as this role’s tendency is not measuring a global explanation; it is measuring an unrelated report card.

Where it stops holding

If the task is always “handle the item in front of you,” a global account is background noise. If the model is retrained weekly, last week’s temperament is already stale, and showing it misleads more than hiding it. Highly personalised models barely have a stable “overall,” and a global sentence will fabricate a character that is not there. This entry does not treat how a few local accounts induce a biased global impression.

Applying it

• Put global copy under “about this system” or in a deployment decision, not beside every result as if it were the reason for this one.
• Write tendency with a scope: “on social-hire résumés in the last three months, files with a gap over six months passed twelve points less often,” not “the system dislikes gaps.”
• Recompute and date it after a version change. Take expired temperament copy down.
• Check: people who have read the global account should be able to answer “how it probably behaves on this class of task,” and should not be able to predict one file’s right or wrong from the same paragraph. If they spend temperament as a verdict, the scope is stuck in the wrong place.

Related

• Same group: L5.02.1 A local explanation accounts for a single output · 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 · L5.03 Trust Calibration · L5.05 The Moderation Principle of Transparency
• Search terms: global explanation · behavioural tendency · partial dependence