The Observation Theory EncyclopediaFrom TSKAboutBy kindBy chapterBy Lean fileLedgerProvenance

attribution instrument

DefinitionAn assignment to each feature of a share of one prediction. The raw gradient is the sensitivity at a row, gradient times input is the share of the change in output, and the averaged squared gradient is a diagonal entry of the read operator, and the three are kept apart. Shapley, permutation, and partial-dependence methods measure other things and are named for them. Chapter 14. Also partial dependence, Shapley.
ExampleFor the affine consumer 3x1 + 2x2, moving from (0, 0) to (1, 1) gives attributions 3 and 2, which sum to the change in output, 5.
BookData Mining as Observation, draft 0.2, commit f3914f0; entry id attribution, kind instrument.
Statusno ledger row names this entry. Corrections: none recorded.
Defining equationnone
Assumptions and scope
  • An assignment to each feature of a share of one prediction. Three gradient quantities are kept apart. The raw gradient is the sensitivity at a row. Gradient times input, the sensitivity times the move along a coordinate, sums exactly to the change in output for an affine consumer and is zero on an unread coordinate. The weighted mean of the squared sensitivity is a diagonal entry of the read operator, so an averaged squared gradient estimates the read operator and a single-row attribution does not.
  • Shapley, permutation, and partial-dependence methods measure other things and are named for them, and for a tree the sensitivity is zero almost everywhere, so importances that count splits are the right reading and finite differences the wrong one.
Prior artnone recorded
Evidencelean/DataMiningAsObservation/Attribution.lean
Reviewedsemantic review 2026-09-06; generated 2026-09-10 from records at the commits on the provenance page.
x1x2x3x4x5
A share of one prediction per feature, and the averaged squared gradient that estimates the read operator.

Equation

none

Conditions

Conditions are curated in entries.toml rather than read from a record.

Ledger

none

First stated

Lundberg and Lee, a unified approach to interpreting model predictions, 2017, as chapter 14 section 14.5 of Data Mining as Observation reads it, with the program’s contraction formula in gtc-prototype/docs/SPECTRUM_FINDINGS.md:90-99.

Measurements

none

Failures and corrections

none

Invariance envelope

none declared

Machine checked

lean/DataMiningAsObservation/Attribution.lean, theorems attr_sum_affine, attr_unread, sq_sensitivity_eq_readOp_diag, at observation-data-mining f3914f0; what the check covers is stated in the book’s appendix C.

Used in

Data Mining as Observation primer L, chapters 0, 8, 14.

Related

sensitivity; read operator; finite difference; consumer.

See also

Book equations stated beside the entry’s terms, not defining it: 0.8, 0.9, 14.5.

Ledger rows that cite the entry’s records without naming it: GO-1, GO-EC-3.

Sources-table rows that share a record with the entry without naming it: chapter 11 section 11.7, chapter 14 section 14.5.

Status

Generated 2026-09-10 by encyclopedia/generate.py; book at observation-data-mining f3914f0; the commit of every record is listed in the encyclopedia’s provenance.

← attribute typeaudit →