About the min-max modeling.

[cjfcsjt]

The research work is exciting! I noticed that you formulate the automated graph contrastive learning as the min-max optimization problem. May I ask that why not use the bi-level optimization where the upper level objective is to search the augmentation operations and the lower level objective is to optimize the contrastive loss (like existing one-shot neural architecture search methods)?

[yyou1996]

Hi @cjfcsjt,

Sorry for the late reply. In paper we did solve it by a naive bi-level optimization if referring to section 3.2.