Hello, the baby data set recall@10 can only reach 0.0553, not 0.0564 in the paper.

[tabo0]

Hello, the baby data set recall@10 can only reach 0.0553, not 0.0564 in the paper.

[enoche]

Hi, thanks for running our code~ May I have a more detailed information about your settings? e.g., Server/Environment/Hyper-parameter ?

[tabo0]

Hello, the code I ran in ubuntu18.4. The GPU is NVIDIA Corporation Device 2204. I'm running main.py without changing the code.

[enoche]

Thanks for your reply.

For your reference, the best hyper-parameters for baby is : ['n_layers', 'reg_weight', 'dropout', 'seed']=(1, 0.1, 0.5, 999), Here is the logs re-generated just now: Baby Logs

Noting different GPU server or seeds may have slightly varied resulted performance which is common in deep learning.

Pls drop me a line if you would like to strictly reproduce my results, I'd glad to print out my evaluation environment for your reference.

Hello, the code I ran in ubuntu18.4. The GPU is NVIDIA Corporation Device 2204. I'm running main.py without changing the code.

[tabo0]

Thank you. Here's my training log. Why is my best result not the highest result of an epoch during training?

[enoche]

Thank you. Here's my training log. Why is my best result not the highest result of an epoch during training?

Good point! In the setting of this repo. best training model is selected as the best performance of validation dataset. Hence, this best model may not be the best in test dataset. While in some other settings, like SimGCL, the best model is selected out from test dataset, which is not applicable in real-world because the test dataset cannot be observed in advance.