Shapley Additive Explanations
A more efficient way to approximate Shapley Values.
Global SHAP
Run SHAP for every instance and obtain a matrix of shapley values.
SHAP Feature Importance, average over all shapley values (for each instance) for a given feature. So we get the average contribution on model output of a feature.
Pros
- all Shapley Values properties
- Model-Agnostic
- “Fair payout” properties
- solid theory
- faster approximation with KernelSHAP
- even faster non Model-Agnostic version with TreeSHAP
- faster computation enables global interpretations
Cons
- all Shapley Values properties
- not selective
- training data access needed
- predictions on unrealistic datapoints if data is correlated
- KernelSHAP still slow
- KernelSHAP ignores dependence between subsets of included and absent features