Classical ML, validation, metrics, interpretation, debugging · insight-0180
State the output scale of SHAP values for boosted classifiers before presenting them
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Data scientist, ML engineer
- Industries
- banking, insurance
Summary: situation, action, reason, boundary
Situation
You show SHAP values from a gradient-boosted classifier as percentage-point changes in probability.
Action
Label plots with the output scale; if stakeholders need probabilities, use the probability output option with interventional background data, or translate examples individually.
Reason
TreeExplainer explains the model's raw output by default, which for binary boosted classifiers is the margin in log-odds.
Boundary
SHAP on XGBoost, LightGBM and other tree ensembles for classification.
Symptom or risk
By default the values are in log-odds, so a 0.5 contribution can mean very different probability changes depending on the baseline.
Exceptions
Regression models' raw output is already on the target scale.
Trade-offs
Probability-scale explanations are intuitive but non-additive in log-odds terms and need a background sample.
Pitfalls
Adding log-odds contributions to a probability baseline in a slide.
Lessons and concepts
Claims and their sources
TreeExplainer's model_output defaults to raw, which is log-odds for XGBoost binary classification; probability output requires interventional feature perturbation with background data.