Reviews: Regret Bounds for Non-decomposable Metrics with Missing Labels

However I understand that space is limited in the paper and in rebuttal, so I'm willing to give the authors the benefit of doubt. Using the same notation for an empirical quantity and its expectation, in a proof about sample complexity, is really too confusing and needs to be fixed before the proof can be verified. Regarding the assumptions, unless I'm missing something, gamma geq b_0 min(b_11, b_01,b_10,b_00) gives zero to some interesting measures. The theoretical discussion is a bit vague about some assumptions and the effect of some parameters of the problem. The experiments compare a variety of data sets but only one competing algorithm.