Explaining classifier outputs with local approximations