In the meantime, the best tricks I've seen for acceptance is to do deanonymization to reviewer #2 and add citations to their work, + add some experiment that #3 requested. And then hope 1 liked it. (Easier said than done).

Nice to see an actual example for using CB's embeddings! (My main problem with them is that you can't get feature importance with embeddings, OR their text features, when using CB's "out of the box"). Did you try other params for the text or embedding features in CB?

Great article! It's one of the clearest i've seen about using IPW practically. Question: In the IPW reweighting, why do you only reweight the "positive" examples (from class 1")? i.e: : 𝚒𝚙_𝚠𝚎𝚒𝚐𝚑𝚝𝚜 = (t / p_hat) + ((𝟷 - t) / (1 - 𝚙_hat)) - Assuming t here is the true class (0/1) , and p_hat - the predicted/prior probability.