722caafb4825ef5d8670710fa29087cf-AuthorFeedback.pdf

Thus, we report the performance (H@1, H@10, MRR)byourimplementation intherighttable. These evidences show that efficiency is not an issue. What is the difference between S1 and S2 in Section 4.2? Why do p2 and p3 behave more differently on42 S1 and behave the same on S2? Reply: Indeed, the difference in AUC-PR score is small. Isitfair tocompareDARTS on training loss with the algorithm optimized based on validation metric?Reply:48 Please note that optimizing architectures on the training set is also an alternative practice in the NAS literature, see49 SNAS[Xieetal.ICLR2019],Auto-DeepLab[Liuetal.CVPR2019],MiLeNAS[Heetal.CVPR2020].50