Completeness

A three-feature explanation may be accurate about those features and still leave most of the prediction unexplained.

Completeness asks how much relevant model behavior the explanation accounts for.

ℹ

Coverage of the model signal

For additive attributions, completeness often requires:

$\sum_i a_i=f(x)-f(x')$

The attributions should reconstruct the difference between the input prediction and baseline. For rules or concepts, completeness needs a task-specific measure of behavior covered.

Balance the Attribution Ledger

Choose how many features to reveal. The visible credits and debits account for more of the prediction difference, while the explanation becomes longer.

Features shown3

Magnitude accounted for

73%

Reader burden

3 line items

FeatureAttribution

income+0.21

debt-0.16

history+0.12

Unshown magnitude0.18

Completeness is an accounting property. A complete ledger can still be a poor interface for a person.

Completeness competes with simplicity. Showing every contributing token can satisfy an accounting identity and bury the decision maker.

Choose the smallest explanation that covers the behavior needed for the user's task, then state what remains outside the explanation.

Checkpoint

Why is a complete explanation not automatically a good explanation?

←PreviousFaithfulnessEvaluating Explanations Next→Stability and RobustnessEvaluating Explanations