KernelSHAP

SHAP for model-agnostic like LIME, but with shapely value guarantees.

1. Suppose $x$ composed of 4 features $f_1,f_2,f_3,f_4$
2. Take coalition samples by replacing excluded features with a background dataset representing them as a coalition vector (0 and 1)
3. Setup a linear regression problem where each datum is the coalition vector and corresponding ML model output Fit an overall model $\hat{y}={\varphi }_0+{\varphi }_1{f}_1+{\varphi }_2{f}_2+{\varphi }_3{f}_3+{\varphi }_4{f}_4$ where ${\varphi }_i$ is the Shapley Values for feature $f_i$