Review for NeurIPS paper: Bandit Samplers for Training Graph Neural Networks

Weaknesses: The authors conduct experiments with 2 layer architecture. However, for few problems and dataset, it may require more complex architecture and the authors could clarify how the proposed algorithm performs in terms of scalability and computation cost. The notation used in the paper sometimes can be confusing. For example - in equation 4 - alpha_ij is a value not a function - alpha_ij(t) can be noted as function of't' It is mentioned that the rewards vary as the training proceeds and it would have been interesting to explore how any of simple bandit algorithms perform in the experiments or how to apply simple bandit algorithms for the current experiments. The authors could try and adapt a simple eps-greedy method to solve this problem The algorithm based on MAB where a combinatorial set of neighbors with size k needs to be chosen.