9cd013fe250ebffc853b386569ab18c0-AuthorFeedback.pdf

Note that the implicit assumption being made is:p(expose = 1,relevance = 1|x) = p(expose = 1|x) 7 p(relevance = 1|x,expose = 1) = p(expose = 1|x) p(relevance = 1|x), so p(relevance = 1|x,expose = 1) =8 p(relevance = 1|x), suggesting thatrelevance exposure|x, i.e. relevance is independent of exposure given the9 features. The other line of prior work is distribution-robust optimization (DRO), which is a vast domain.