#local-fidelity

Question 1: Single choice

LIME weights each perturbed sample $z$ by $\pi_x(z) = \exp(-D(x, z)^2 / \sigma^2)$ before fitting its linear surrogate. What is the main risk of choosing a very large kernel width $\sigma$ ?

Distant samples count almost fully, so the fit can miss or flip the local effect1
All samples get weights close to zero, so the surrogate has almost no data left to fit2
The surrogate becomes a nonlinear model, which makes the explanation harder to read3
The black-box model has to be retrained on the newly weighted perturbed samples4

Timed test