Interpretable Machine Learning
Explain model predictions.
Global vs Local
Global: overall feature importance. Local: individual prediction explanation.
Methods
SHAP: game-theoretic, consistent. LIME: local linear approximation. ICE: individual conditional expectations.
Model-Specific
Linear models: coefficients. Decision trees: rules. Neural networks: attention visualization.
Key Takeaways
- Global vs local explanation scope
- SHAP provides consistent attributions
- Model-specific methods are simpler