This paper is quite borderline. The reviewers were all learning towards acceptance, but none were willing to fight for the paper. The main concern for the paper is that the empirical performance gain over Kipf et al is quite small, and in some cases, non-existent. In the balance, I would put this paper slightly over the bar for acceptance. However, we strongly encourage the authors to try to better highlight the benefits over Kipf et al regarding unseen nodes, in the final version. We also encourage the authors to quantitatively evaluate the speed difference from Alet et al, e.g. for one iteration of the algorithm, and to address the points raised by the reviewers.