#lime

Question 1: Single choice

How does LIME explain the prediction of a black-box classifier for one input $x$ ?

It fits a proximity-weighted simple model to the outputs on perturbed copies of $x$ 1
It back-propagates the class score and uses the input gradient at $x$ as the explanation2
It averages each feature's marginal contribution over all possible orderings of the features3
It trains one global decision tree that imitates the black box on the whole training set4