There are several papers dealing with the issue of sampling when training a graph neural network, and this paper provides a new approach to it using (adversarial) bandit technology. This idea makes sense and the authors provide convincing arguments for why this technique should improve upon previous art. Although there may be room for slightly more comprehensive experiments and reasoning for the details of the proposed solution, the paper provides a clear representation of a novel method that works well for an important problem. This paper would make a good addition to NeurIPS.