Feature Importance and Explanations¶
MPSBoost provides multiple explanation layers:
- gain importance
- split-count importance
- permutation importance
- controlled SHAP-like approximate explanations
Design principles¶
- Do not duplicate prediction or scoring logic.
- Do not present approximate explanations as official SHAP.
- Advance the official SHAP TreeExplainer adapter as a separate task.
Usage direction¶
When research or reporting needs official SHAP semantics, wait for the S16.5a / S23 adapter
documentation. For the current package, use permutation_importance as model-agnostic
interpretation.
Official SHAP path¶
Official SHAP is optional and explicit:
python -m pip install 'mpsboost[shap]'
export_native_trees_for_shap(estimator) exports native tree structure for TreeExplainer adapter
validation without training data, credentials, telemetry, or device identifiers.
official_shap_tree_explainer(...) stops clearly until semantic validation is enabled, so
approximate explanations are not presented as official SHAP output.