Review for NeurIPS paper: Graph Information Bottleneck

Specifically: (1) The author mentioned some information-related graph representation works, such as Deep Graph Infomax, InfoGraph, etc, which seems to have similar intuition and technical implementation with this work. However, this paper only briefly discuss that these works are for unsupervised learning, while this paper focused on robust supervised training. This is not enough to clearly state the contribution of this work's approach. Better to elaborate more on the technical difference. GPT-GNN: Generative Pre-Training of Graph Neural Networks (https://arxiv.org/abs/2006.15437), it would be better that the authors can consider adding them as comparison baselines, as the proposed approach and these papers all leverage additional signal to regularize training. Update: Overall, the authors' rebuttal solve many of my concerns, and I do think it has enough novelty and contribution in this field, so I decide to raise up my score.