tginart / deletion-efficient-kmeans Goto Github PK

Hi @tginart,

Thanks for making your code public. I'm very interested in your work and would like to understand the whole procedure better. However, I have a question about the gamma-imbalanced correction step. How can we guarantee the cluster size is at least self.momentum after applying self._momentum_correction? I could not find proof of this in the paper and it is hard for me to imagine how we can guarantee the cluster size by averaging two centroids by weights. Moreover, the momentum correction step is sequential and the later correction may have influenced previous corrections, which makes the theoretical analysis even harder.

Please let me know if I misunderstand something here. And thanks for your response in advance!

Best,
Chao