Ce Integration Compare

CE Integration Compare

You are helping users work with the CE integrations for SHAP and LIME.

Required references

• src/calibrated_explanations/integrations/shap.py
• src/calibrated_explanations/integrations/lime.py
• src/calibrated_explanations/integrations/__init__.py

Use this skill when

• A user asks how to use CE with SHAP or LIME.
• Comparing CE explanations with SHAP/LIME outputs.
• Configuring or debugging the integration adapters.
• Understanding the differences between calibrated explanations and feature-attribution methods.

Key concepts

CE vs. feature attribution methods

Calibrated explanations provide rule-based explanations with calibrated uncertainty intervals. SHAP and LIME provide feature attribution values (importance scores per feature). The approaches are complementary:

• CE rules tell you which conditions justify a prediction with calibrated confidence.
• SHAP/LIME tell you how much each feature contributed to the prediction.

Integration architecture

The integration modules provide adapter classes that:

1. Wrap a fitted CE explainer to produce SHAP-compatible or LIME-compatible outputs.
2. Allow side-by-side comparison of CE rules with feature attributions.
3. Preserve CE calibration semantics while exposing familiar interfaces.

Workflow

1. Fit and calibrate a CE pipeline first (use ce-pipeline-builder).
2. Import the integration adapter:

   from calibrated_explanations.integrations.shap import CESHAPAdapter
   # or
   from calibrated_explanations.integrations.lime import CELIMEAdapter
   

3. Generate comparison explanations using both CE and the adapter.
4. Interpret the results: CE rules for conditions, SHAP/LIME values for attribution magnitudes.

Constraints

• Always fit and calibrate the CE pipeline before using integration adapters.
• Integration adapters are convenience wrappers; the underlying CE explainer must be properly configured.
• Do not use integration adapters as replacements for CE explanations; they are comparison tools.