Skip to main content

Accumulated local effects

Inspect accumulated local changes in predictions across a tabular feature.

Join waitlist
Testing group
Explainability
Applies to
Tabular models
App version
2.0.0
Channel
stable

Required inputs

A compatible model and representative inputs; some explanation and faithfulness methods also require labels or task-specific reference regions.

Test procedure

Generate the explanation with recorded method settings. Inspect representative and difficult cases, and use an applicable faithfulness check where available.

From local differences to feature effects

  1. Partition the feature
  2. Measure local differences
  3. Accumulate
  4. Center the effects
Partition the selected feature into bins, measure prediction differences within occupied bins, then accumulate and center the local effects.

Interpreting results

Use the explanation to investigate which inputs influence a prediction and whether that account remains consistent with the model's behaviour.

Limitations

Evidence record

Preserve the input, prediction, explanation, method configuration and any faithfulness results used in the review.