Classical ML, validation, metrics, interpretation, debugging · insight-1151
State the background data behind SHAP values, because it defines what a feature value is compared with
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- banking, insurance, cross-industry
Summary: situation, action, reason, boundary
Situation
SHAP bars for a tree model are presented as how much each feature matters, with no mention of the reference rows.
Action
Record feature_perturbation and the background sample (size and how it was drawn), choose the model output scale deliberately, and recompute with a second background to check stability.
Reason
SHAP values plus the expected value equal the model output; the expected value is the average prediction over the background, so the reference defines the baseline.
Boundary
TreeExplainer on boosted trees and forests in reports, audits and dashboards.
Symptom or risk
Attributions are differences from an expected value computed over a reference distribution; a different background, or the training-leaf counts in path-dependent mode, changes the numbers and the story.
Exceptions
For a purely relative ranking of features in one fixed analysis, conclusions are often stable across reasonable backgrounds, which you should verify.
Trade-offs
Interventional mode needs a background and more compute; path-dependent needs none but uses training leaf counts.
Pitfalls
Comparing SHAP values computed with different backgrounds across model versions.
Lessons and concepts
Claims and their sources
TreeExplainer's interventional mode requires a background dataset (100 to 1000 random samples suggested); tree_path_dependent uses per-leaf training counts as the background; SHAP values plus expected_value equal the model output in the chosen output space.